Frontier LLM API providers

Q: What actually happens when a model I depend on is deprecated?

- The alias stops resolving and requests 404 - A pinned dated snapshot keeps serving until its published retirement date - A bare alias means you migrate today - Anthropic and OpenAI publish retirement dates months ahead; several providers publish nothing - Rolling aliases can change behaviour with no version bump

Q: Is DeepSeek or Kimi really good enough to replace Claude or GPT?

- For extraction, classification, summarisation, translation and bulk rewriting: yes - Roughly 1/35th the output cost of Anthropic's volume model, or 1/120th of its flagship - The honest bulk-work comparison is against the volume tier, not the flagship - For long agentic runs, sustained instruction following and first-try code, frontier labs stay ahead - Split the workload rather than picking one provider for everything

Q: Does an OpenAI-compatible endpoint actually make switching easy?

- It makes the transport identical and the semantics different - Chat completions port cleanly - Tool-call formats, effort parameters, cache accounting, thinking blocks and schema enforcement do not - Prompts tuned on one model regress on another - Budget a re-evaluation, not a config change

Q: How much does prompt caching really save on an agent loop?

- More than any other lever - Cache reads typically bill at 10-25% of the input rate - An agent loop resends the whole conversation each turn, so a long run is 80-90% cache reads - The trap: caching is a prefix match, so a timestamp or reordered tool list invalidates everything after it

Q: Which providers can serve inference from an Australian region?

- Amazon Bedrock (ap-southeast-2) and Google Vertex AI (australia-southeast1) serve frontier models in-country - First-party Anthropic, OpenAI and xAI endpoints route to US or EU by default - Residency controls vary by plan - Self-hosting is the only unambiguously AU-resident answer

Q: Should I go through a router like OpenRouter instead of direct?

- Use a router when model choice is a runtime decision, for outage fallback, or while benchmarking - Go direct when you depend on provider-specific features - That means cache breakpoints, computer use, batch endpoints, enterprise retention terms - Routers expose the intersection of upstream features, not the union

Q: Why is the time-to-first-token column mostly empty?

- Almost nobody publishes it - Self-reported latency is meaningless without a stated prompt length, region and concurrency - Figures shown are community measurements for speed-focused providers - toolweight will fill it only after running its own harness with published methodology

Anthropic, OpenAI and Google lead on capability, cheaper Chinese labs win on price, routers trade features for reach; pick on continuity, not headline price.

Providers: 18
Fields compared: 27
Source confidence: 34%
Last verified: 2026-07-30 (1mo ago)
Re-verified: every 7 days

Data confidence0% of 369 figures verified in the last 7d

142sourced of 369330stale, oldest 8mo ago

All comparisons

Comparison	Tools	Verified
Sandbox providers	20	2026-07-23
Coding agents	17	2026-01-15
LLM APIs	18	2026-07-30
Browser APIs	13	2026-07-01
S3-compatible storage	12	2026-07-23
Vercel alternatives	13	2026-06-15
Convex alternatives	11	2026-07-20
Email APIs	13	2026-06-05
Database engines	5	2026-07-01
Domain registrars	15	2026-06-30
Domain APIs	15	2026-07-23
AU SMS APIs	13	2026-07-20

Providers

18 compared

Google GeminiWorkloads that genuinely need 500K+ tokens of context per request

76.9score

Amazon BedrockRegulated workloads that must stay inside an existing AWS perimeter

75.6score

Mistral AIEU data-residency requirements that rule out US-hosted inference

75.2score

OpenAIMultimodal products that need speech and image generation alongside text

73.2score

QwenFinding an open-weight model at a specific size or modality

68.6score

xAIProducts that want real-time X/Twitter context in the model

64.5score

Meta LlamaWorkloads that must survive any single vendor disappearing

64.3score

Self-hosted (vLLM)Data that legally cannot leave your infrastructure

63.8score

CohereRAG pipelines where rerank quality drives the result

63.6score

Together AIRunning DeepSeek or Kimi weights under a Western contract

62.9score

Fireworks AIOpen-weight workloads where output must validate against a schema every time

61.9score

AnthropicAgent harnesses that call tools for minutes at a time

61.0score

Moonshot AIAgent loops on an open-weight budget

60.0score

OpenRouterKeeping model choice a runtime decision rather than an architectural one

59.6score

Z.ai (GLM)Driving a coding agent all day on a predictable monthly cost

58.2score

GroqVoice agents and any interface where perceived latency is the feature

55.3score

CerebrasReasoning models where thinking-token latency is the bottleneck

55.2score

DeepSeekHigh-volume classification, extraction and summarisation

53.3score

Differences only

18 tools · verified 1mo ago

Frontier LLM API providers, 18 tools compared across 27 of 27 fields.
Tool	Weighted
			w5	w3	w4	w3	w4	w5	w4	w4	w8	w9	w6	w4	w8	w7	w6	w3	w5			w7	w7	w6	w7	w4	w1
Google GeminiGemini via AI Studio for prototyping or Vertex AI for production	76.9	Gemini 3.1 Pro Preview (gemini-3.1-pro-preview)	1,048,576tokens	65,536tokens	Yes	Partial	Partial	Yes	·	·	$2/M tok	$12/M tok	$0.2/M tok	50%	Yes	Yes	Yes	Partial	Partial	Explicit caches, default 1 h TTL; implicit caching too	Stable versions carry dated suffixes you can pin, and Google publishes a retirement date per version. In practice this is the fastest-moving line-up here: preview models are withdrawn with weeks of notice, and the 1.5 generation was retired roughly a year after GA. Vertex adds contractual predictability that AI Studio does not.	–	Yes	Partial	Partial	Yes	2023	Long context and native multimodality across Google's cloud
Amazon BedrockMulti-vendor model access inside your existing AWS account	75.6	Multi-vendor, Claude Opus 4.8, Llama 4, Mistral, Nova Premier	1,000,000tokens	128,000tokens	Yes	Partial	Partial	No	·	·	–	·	·	50%	Yes	Yes	Yes	Yes	No	5 min default, 1 h option; explicit breakpoints only	The most contractual answer in the category. Bedrock model IDs are versioned and move through a published Active to Legacy to end-of-life lifecycle with dated notices in the console and documentation, and legacy versions keep serving existing workloads after a successor lands. The cost of that stability is lag: new features reach Bedrock weeks to months after the first-party API, and some never do.	–	Yes	Yes	No	Yes	2023	Frontier models inside your existing AWS security perimeter
Mistral AIEuropean lab with an open-weight lineage and EU-resident hosting	75.2	Mistral Large 3 (mistral-large-3-25-12)	256,000tokens	–	Yes	Partial	Yes	Yes	·	·	$0.5/M tok	$1.5/M tok	$0.05/M tok	50%	Yes	Yes	Partial	No	Partial	·	Dated model IDs (the -2411 style suffix) are stable and documented, and Mistral maintains a legacy-model page listing deprecation and retirement dates. The stronger guarantee is structural: for the Apache-licensed models you can keep serving the same weights yourself after the endpoint goes away.	–	Yes	Partial	No	No	2023	European models with open weights and deployable anywhere
OpenAIGPT models plus audio, images and embeddings on one bill	73.2	GPT-5.6 Sol (gpt-5.6-sol)	1,050,000tokens	128,000tokens	Yes	Partial	Partial	Yes	·	·	$5/M tok	$30/M tok	$0.5/M tok	50%	Yes	Yes	Yes	Partial	Yes	Automatic, ≥30 min, no configuration	Dated snapshots are the norm and remain callable after the alias advances, which is the strongest routine protection in the category. Retirements are announced on a public deprecations page, but the notice window has historically been shorter than Anthropic's, closer to a quarter than half a year for API snapshots, and preview models have been pulled faster still.	90days	Yes	Partial	No	Unknown	2020	One API for text, reasoning, audio, images and embeddings
QwenAlibaba's model family, huge open-weight range, closed flagship	68.6	Qwen3.7-Max (qwen3.7-max)	1,000,000tokens	·	Partial	Partial	Partial	Yes	·	·	$2.5/M tok	$7.5/M tok	$0.25/M tok	50%	Yes	Partial	Partial	No	Partial	Explicit cache 5 min (extends on hit); implicit cache no fixed TTL	Model Studio exposes dated snapshot IDs alongside rolling aliases, so pinning is possible if you use the dated form. Deprecation notices appear in the console rather than as a public policy page, and the cadence of new Qwen releases means aliases move quickly. Open-weight tiers remove the problem entirely.	·	Yes	Unknown	Partial	Partial	2023	The widest open-weight family, plus a closed flagship tier
xAIGrok models with an OpenAI-shaped API and live X data access	64.5	Grok 4.5 (grok-4.5)	500,000tokens	–	Yes	Partial	Partial	Yes	·	·	$2/M tok	$6/M tok	$0.3/M tok	–	Yes	Yes	Partial	No	No	·	Dated model IDs exist and are the recommended way to address a model, but xAI publishes no formal deprecation policy or notice window, and model naming has changed repeatedly between generations. The shortest track record of any first-party lab here.	–	Yes	Unknown	Partial	No	2024	Fast, cheap frontier models with live access to X
Meta LlamaOpen-weight Llama models, hosted almost everywhere but Meta	64.3	Llama 4 Maverick	1,000,000tokens	–	Yes	No	Yes	Yes	·	·	–	·	·	·	Partial	Partial	No	No	No	Host-dependent	Effectively unlimited, and the reason Llama matters strategically. The weights are downloadable and mirrored, so no vendor can retire the model out from under you, if a host drops it, three others still serve it, or you run it yourself. Meta may stop releasing new generations, but nothing already published disappears.	–	Yes	Partial	No	Yes	2023	Open-weight models you can host anywhere, forever
Self-hosted (vLLM)baseline · Run open weights on your own GPUs behind an OpenAI-shaped API	63.8	Whatever you deploy, DeepSeek-V3.x, Qwen3, Kimi K2, Llama 4	–	–	Partial	No	Yes	Yes	–	·	–	·	–	0%	Partial	Yes	No	No	No	Automatic prefix cache in VRAM, no TTL	Permanent, and this is the whole point of the baseline. The checkpoint sits on your disk. Nobody can deprecate it, re-price it, silently upgrade it, or change its behaviour between releases. The obligation transfers to you: you own the GPUs, the upgrades, the security patches and the on-call roster.	–	Yes	Yes	No	Yes	2023	Your weights, your GPUs, your OpenAI-compatible endpoint
CohereEnterprise-focused models built for RAG and private deployment	63.6	Command A+ (command-a-plus-05-2026)	128,000tokens	64,000tokens	Yes	No	Yes	Yes	·	·	–	–	·	·	Yes	Yes	No	No	No	·	Dated model names are the default (command-a-03-2025), and Cohere publishes deprecation notices with replacement guidance. The strongest guarantee is deployment-shaped rather than policy-shaped: a private VPC or on-premise deployment continues running whatever version you licensed regardless of what the hosted platform does.	·	Yes	Partial	No	Partial	2021	Enterprise RAG models you can deploy inside your own network
Together AIServerless and dedicated hosting for open-weight models	62.9	Open-weight catalogue, DeepSeek-V3.x, Kimi K2, Qwen3, Llama 4	–	–	Partial	Partial	Yes	Yes	·	·	–	·	·	50%	Partial	Yes	No	No	No	·	Serverless models are rotated off when demand drops, typically with a deprecations page and short notice measured in weeks. Because everything served is open-weight, the mitigation is always available: move to a dedicated endpoint on the same platform, to another host, or onto your own GPUs with the same checkpoint.	–	Yes	Unknown	No	No	2023	Open-weight models with a Western contract and dedicated capacity
Fireworks AIFast open-weight inference with strong structured-output support	61.9	Open-weight catalogue, DeepSeek, Kimi K2, Qwen3, Llama 4	–	–	Partial	Partial	Yes	Yes	·	·	–	·	·	·	Partial	Yes	No	No	No	·	Same shape as Together: a published deprecations list, serverless models retired on weeks of notice when demand falls, and dedicated deployments as the stable option. Open weights throughout mean no retirement is terminal, only inconvenient.	·	Yes	Unknown	No	No	2023	Low-latency open-weight inference with strict structured output
AnthropicClaude models, built around long agentic runs and tool use	61.0	Claude Fable 5 (claude-fable-5)	1,000,000tokens	128,000tokens	Yes	No	No	Partial	·	·	$10/M tok	$50/M tok	$1/M tok	50%	Yes	Yes	Yes	Yes	Yes	5 min default, 1 h option; automatic or explicit breakpoints	Published deprecation page lists a retirement date per model, typically months ahead, Opus 3 went on 2026-01-05, Sonnet 3.7 and Haiku 3.5 on 2026-02-19. Older models carried dated snapshot IDs you could pin, but the current generation ships as bare aliases with no dated ID behind them, so the retirement date is your only guarantee.	180days	Partial	Partial	No	Partial	2023	Frontier models for long-horizon agentic work and code
Moonshot AIKimi models, open-weight agentic performance at low cost	60.0	Kimi K3 (kimi-k3)	1,048,576tokens	·	Yes	No	Yes	Yes	·	–	$3/M tok	$15/M tok	$0.3/M tok	–	Yes	Partial	Partial	No	No	Automatic, no fixed TTL	Dated model IDs (the -0905 style suffix) are used and remain addressable, which is better than DeepSeek's rolling aliases. No published deprecation policy or notice window. As with the other open-weight labs, the real guarantee is the licence, download the checkpoint and continuity is your problem, not theirs.	·	Yes	No	Partial	No	2023	Open-weight agentic performance at open-weight prices
OpenRouterOne OpenAI-shaped key in front of hundreds of models	59.6	Whichever upstream model you route to (400+ available)	–	–	Partial	Partial	Partial	Yes	–	·	–	·	·	0%	Yes	Partial	Partial	No	No	Passed through where the upstream supports caching	Structurally the best hedge on this page and formally the weakest promise. OpenRouter guarantees nothing itself, a closed model vanishes the moment its lab retires it. But for open-weight models the same ID stays routable because several independent hosts serve it, and automatic failover means a single host going dark is invisible to your application. Buy it as insurance against host failure, not against model retirement.	–	Partial	Partial	No	No	2023	One key, one API shape, every model worth calling
Z.ai (GLM)GLM models with an unusually cheap flat-rate coding plan	58.2	GLM-4.6	200,000tokens	·	Partial	No	Yes	Yes	·	·	$0.6/M tok	$2.2/M tok	·	·	Yes	Partial	Partial	No	No	·	Versioned model names (GLM-4.5, 4.6) stay addressable after a successor ships, but there is no published deprecation policy or notice period. MIT-licensed weights are the fallback: every flagship generation so far has been released publicly, so a retired endpoint does not strand the model.	·	Yes	No	Partial	No	2023	Coding-focused GLM models with a flat-rate subscription
GroqCustom LPU silicon serving open-weight models at extreme speed	55.3	Curated open-weight models (Kimi K2, Llama 4, Qwen3, GPT-OSS)	–	–	Partial	Partial	Yes	Yes	250ms	400tok/s	–	·	·	–	Partial	Partial	No	No	No	·	The weakest continuity of any host here. Models are added and removed from the console on a short cycle as Groq re-allocates LPU capacity, deprecations are announced with weeks rather than months of notice, and there is no dedicated-endpoint escape hatch for a model that gets pulled. Design for substitution: keep two model IDs configured and an eval you can rerun.	–	Partial	Unknown	No	No	2024	Custom LPU silicon for open-weight models at very low latency
CerebrasWafer-scale inference, the fastest tokens per second available	55.2	Curated open-weight models (Qwen3, GLM, Llama, GPT-OSS)	–	–	No	No	Yes	Yes	200ms	2,000tok/s	–	·	·	·	Partial	Partial	No	No	No	·	Same fragility as Groq, from the same cause: a small curated menu sized to available wafer-scale capacity, with models added and removed as demand shifts. No published notice window. The models themselves are open-weight, so a removal costs you a re-benchmark and a host change rather than a rewrite.	·	Partial	Unknown	No	No	2024	Wafer-scale inference, the highest tokens per second available
DeepSeekFrontier-adjacent models at a small fraction of Western prices	53.3	DeepSeek-V4-Flash (deepseek-v4-flash)	1,000,000tokens	384,000tokens	No	No	Yes	Yes	·	·	$0.14/M tok	$0.28/M tok	$0.0028/M tok	0%	Yes	Partial	Partial	No	No	Automatic disk cache, no fixed TTL	The worst on this page for a closed endpoint, and the best if you self-host. The API offers only rolling aliases, deepseek-chat silently points at whatever the current model is, so behaviour can change with no version to pin and no deprecation notice. The mitigation is the MIT licence: download the checkpoint and the version is yours permanently.	–	No	No	Yes	No	2023	Near-frontier quality at a fraction of the token price

Google GeminiGemini via AI Studio for prototyping or Vertex AI for production76.9

Capability

Flagship model: Gemini 3.1 Pro Preview (gemini-3.1-pro-preview)
Context window (tokens): 1,048,576tokens
Max output (tokens): 65,536tokens
Image input: Yes
Audio in/out: Partial
Open weights: Partial
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $2/M tok
$/M output (/M tok): $12/M tok
$/M cache read (/M tok): $0.2/M tok
Batch discount (%): 50%

Agentic

Tool use: Yes
Schema output: Yes
Effort control: Yes
Computer use: Partial
MCP support: Partial
Cache TTL: Explicit caches, default 1 h TTL; implicit caching too

Governance & continuity

Continuity policy: Stable versions carry dated suffixes you can pin, and Google publishes a retirement date per version. In practice this is the fastest-moving line-up here: preview models are withdrawn with weeks of notice, and the 1.5 generation was retired roughly a year after GA. Vertex adds contractual predictability that AI Studio does not.
Notice period (days): –
Pinnable versions: Yes
Zero retention: Partial
Trains on your data: Partial
AU region: Yes

Traction

API since: 2023

Positioning

Positioning: Long context and native multimodality across Google's cloud

Amazon BedrockMulti-vendor model access inside your existing AWS account75.6

Capability

Flagship model: Multi-vendor, Claude Opus 4.8, Llama 4, Mistral, Nova Premier
Context window (tokens): 1,000,000tokens
Max output (tokens): 128,000tokens
Image input: Yes
Audio in/out: Partial
Open weights: Partial
OpenAI-compat API: No

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): ·
Batch discount (%): 50%

Agentic

Tool use: Yes
Schema output: Yes
Effort control: Yes
Computer use: Yes
MCP support: No
Cache TTL: 5 min default, 1 h option; explicit breakpoints only

Governance & continuity

Continuity policy: The most contractual answer in the category. Bedrock model IDs are versioned and move through a published Active to Legacy to end-of-life lifecycle with dated notices in the console and documentation, and legacy versions keep serving existing workloads after a successor lands. The cost of that stability is lag: new features reach Bedrock weeks to months after the first-party API, and some never do.
Notice period (days): –
Pinnable versions: Yes
Zero retention: Yes
Trains on your data: No
AU region: Yes

Traction

API since: 2023

Positioning

Positioning: Frontier models inside your existing AWS security perimeter

Mistral AIEuropean lab with an open-weight lineage and EU-resident hosting75.2

Capability

Flagship model: Mistral Large 3 (mistral-large-3-25-12)
Context window (tokens): 256,000tokens
Max output (tokens): –
Image input: Yes
Audio in/out: Partial
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $0.5/M tok
$/M output (/M tok): $1.5/M tok
$/M cache read (/M tok): $0.05/M tok
Batch discount (%): 50%

Agentic

Tool use: Yes
Schema output: Yes
Effort control: Partial
Computer use: No
MCP support: Partial
Cache TTL: ·

Governance & continuity

Continuity policy: Dated model IDs (the -2411 style suffix) are stable and documented, and Mistral maintains a legacy-model page listing deprecation and retirement dates. The stronger guarantee is structural: for the Apache-licensed models you can keep serving the same weights yourself after the endpoint goes away.
Notice period (days): –
Pinnable versions: Yes
Zero retention: Partial
Trains on your data: No
AU region: No

Traction

API since: 2023

Positioning

Positioning: European models with open weights and deployable anywhere

OpenAIGPT models plus audio, images and embeddings on one bill73.2

Capability

Flagship model: GPT-5.6 Sol (gpt-5.6-sol)
Context window (tokens): 1,050,000tokens
Max output (tokens): 128,000tokens
Image input: Yes
Audio in/out: Partial
Open weights: Partial
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $5/M tok
$/M output (/M tok): $30/M tok
$/M cache read (/M tok): $0.5/M tok
Batch discount (%): 50%

Agentic

Tool use: Yes
Schema output: Yes
Effort control: Yes
Computer use: Partial
MCP support: Yes
Cache TTL: Automatic, ≥30 min, no configuration

Governance & continuity

Continuity policy: Dated snapshots are the norm and remain callable after the alias advances, which is the strongest routine protection in the category. Retirements are announced on a public deprecations page, but the notice window has historically been shorter than Anthropic's, closer to a quarter than half a year for API snapshots, and preview models have been pulled faster still.
Notice period (days): 90days
Pinnable versions: Yes
Zero retention: Partial
Trains on your data: No
AU region: Unknown

Traction

API since: 2020

Positioning

Positioning: One API for text, reasoning, audio, images and embeddings

QwenAlibaba's model family, huge open-weight range, closed flagship68.6

Capability

Flagship model: Qwen3.7-Max (qwen3.7-max)
Context window (tokens): 1,000,000tokens
Max output (tokens): ·
Image input: Partial
Audio in/out: Partial
Open weights: Partial
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $2.5/M tok
$/M output (/M tok): $7.5/M tok
$/M cache read (/M tok): $0.25/M tok
Batch discount (%): 50%

Agentic

Tool use: Yes
Schema output: Partial
Effort control: Partial
Computer use: No
MCP support: Partial
Cache TTL: Explicit cache 5 min (extends on hit); implicit cache no fixed TTL

Governance & continuity

Continuity policy: Model Studio exposes dated snapshot IDs alongside rolling aliases, so pinning is possible if you use the dated form. Deprecation notices appear in the console rather than as a public policy page, and the cadence of new Qwen releases means aliases move quickly. Open-weight tiers remove the problem entirely.
Notice period (days): ·
Pinnable versions: Yes
Zero retention: Unknown
Trains on your data: Partial
AU region: Partial

Traction

API since: 2023

Positioning

Positioning: The widest open-weight family, plus a closed flagship tier

xAIGrok models with an OpenAI-shaped API and live X data access64.5

Capability

Flagship model: Grok 4.5 (grok-4.5)
Context window (tokens): 500,000tokens
Max output (tokens): –
Image input: Yes
Audio in/out: Partial
Open weights: Partial
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $2/M tok
$/M output (/M tok): $6/M tok
$/M cache read (/M tok): $0.3/M tok
Batch discount (%): –

Agentic

Tool use: Yes
Schema output: Yes
Effort control: Partial
Computer use: No
MCP support: No
Cache TTL: ·

Governance & continuity

Continuity policy: Dated model IDs exist and are the recommended way to address a model, but xAI publishes no formal deprecation policy or notice window, and model naming has changed repeatedly between generations. The shortest track record of any first-party lab here.
Notice period (days): –
Pinnable versions: Yes
Zero retention: Unknown
Trains on your data: Partial
AU region: No

Traction

API since: 2024

Positioning

Positioning: Fast, cheap frontier models with live access to X

Meta LlamaOpen-weight Llama models, hosted almost everywhere but Meta64.3

Capability

Flagship model: Llama 4 Maverick
Context window (tokens): 1,000,000tokens
Max output (tokens): –
Image input: Yes
Audio in/out: No
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): ·
Batch discount (%): ·

Agentic

Tool use: Partial
Schema output: Partial
Effort control: No
Computer use: No
MCP support: No
Cache TTL: Host-dependent

Governance & continuity

Continuity policy: Effectively unlimited, and the reason Llama matters strategically. The weights are downloadable and mirrored, so no vendor can retire the model out from under you, if a host drops it, three others still serve it, or you run it yourself. Meta may stop releasing new generations, but nothing already published disappears.
Notice period (days): –
Pinnable versions: Yes
Zero retention: Partial
Trains on your data: No
AU region: Yes

Traction

API since: 2023

Positioning

Positioning: Open-weight models you can host anywhere, forever

Self-hosted (vLLM)Run open weights on your own GPUs behind an OpenAI-shaped API63.8

Capability

Flagship model: Whatever you deploy, DeepSeek-V3.x, Qwen3, Kimi K2, Llama 4
Context window (tokens): –
Max output (tokens): –
Image input: Partial
Audio in/out: No
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: –
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): –
Batch discount (%): 0%

Agentic

Tool use: Partial
Schema output: Yes
Effort control: No
Computer use: No
MCP support: No
Cache TTL: Automatic prefix cache in VRAM, no TTL

Governance & continuity

Continuity policy: Permanent, and this is the whole point of the baseline. The checkpoint sits on your disk. Nobody can deprecate it, re-price it, silently upgrade it, or change its behaviour between releases. The obligation transfers to you: you own the GPUs, the upgrades, the security patches and the on-call roster.
Notice period (days): –
Pinnable versions: Yes
Zero retention: Yes
Trains on your data: No
AU region: Yes

Traction

API since: 2023

Positioning

Positioning: Your weights, your GPUs, your OpenAI-compatible endpoint

CohereEnterprise-focused models built for RAG and private deployment63.6

Capability

Flagship model: Command A+ (command-a-plus-05-2026)
Context window (tokens): 128,000tokens
Max output (tokens): 64,000tokens
Image input: Yes
Audio in/out: No
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): –
$/M output (/M tok): –
$/M cache read (/M tok): ·
Batch discount (%): ·

Agentic

Tool use: Yes
Schema output: Yes
Effort control: No
Computer use: No
MCP support: No
Cache TTL: ·

Governance & continuity

Continuity policy: Dated model names are the default (command-a-03-2025), and Cohere publishes deprecation notices with replacement guidance. The strongest guarantee is deployment-shaped rather than policy-shaped: a private VPC or on-premise deployment continues running whatever version you licensed regardless of what the hosted platform does.
Notice period (days): ·
Pinnable versions: Yes
Zero retention: Partial
Trains on your data: No
AU region: Partial

Traction

API since: 2021

Positioning

Positioning: Enterprise RAG models you can deploy inside your own network

Together AIServerless and dedicated hosting for open-weight models62.9

Capability

Flagship model: Open-weight catalogue, DeepSeek-V3.x, Kimi K2, Qwen3, Llama 4
Context window (tokens): –
Max output (tokens): –
Image input: Partial
Audio in/out: Partial
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): ·
Batch discount (%): 50%

Agentic

Tool use: Partial
Schema output: Yes
Effort control: No
Computer use: No
MCP support: No
Cache TTL: ·

Governance & continuity

Continuity policy: Serverless models are rotated off when demand drops, typically with a deprecations page and short notice measured in weeks. Because everything served is open-weight, the mitigation is always available: move to a dedicated endpoint on the same platform, to another host, or onto your own GPUs with the same checkpoint.
Notice period (days): –
Pinnable versions: Yes
Zero retention: Unknown
Trains on your data: No
AU region: No

Traction

API since: 2023

Positioning

Positioning: Open-weight models with a Western contract and dedicated capacity

Fireworks AIFast open-weight inference with strong structured-output support61.9

Capability

Flagship model: Open-weight catalogue, DeepSeek, Kimi K2, Qwen3, Llama 4
Context window (tokens): –
Max output (tokens): –
Image input: Partial
Audio in/out: Partial
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): ·
Batch discount (%): ·

Agentic

Tool use: Partial
Schema output: Yes
Effort control: No
Computer use: No
MCP support: No
Cache TTL: ·

Governance & continuity

Continuity policy: Same shape as Together: a published deprecations list, serverless models retired on weeks of notice when demand falls, and dedicated deployments as the stable option. Open weights throughout mean no retirement is terminal, only inconvenient.
Notice period (days): ·
Pinnable versions: Yes
Zero retention: Unknown
Trains on your data: No
AU region: No

Traction

API since: 2023

Positioning

Positioning: Low-latency open-weight inference with strict structured output

AnthropicClaude models, built around long agentic runs and tool use61.0

Capability

Flagship model: Claude Fable 5 (claude-fable-5)
Context window (tokens): 1,000,000tokens
Max output (tokens): 128,000tokens
Image input: Yes
Audio in/out: No
Open weights: No
OpenAI-compat API: Partial

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $10/M tok
$/M output (/M tok): $50/M tok
$/M cache read (/M tok): $1/M tok
Batch discount (%): 50%

Agentic

Tool use: Yes
Schema output: Yes
Effort control: Yes
Computer use: Yes
MCP support: Yes
Cache TTL: 5 min default, 1 h option; automatic or explicit breakpoints

Governance & continuity

Continuity policy: Published deprecation page lists a retirement date per model, typically months ahead, Opus 3 went on 2026-01-05, Sonnet 3.7 and Haiku 3.5 on 2026-02-19. Older models carried dated snapshot IDs you could pin, but the current generation ships as bare aliases with no dated ID behind them, so the retirement date is your only guarantee.
Notice period (days): 180days
Pinnable versions: Partial
Zero retention: Partial
Trains on your data: No
AU region: Partial

Traction

API since: 2023

Positioning

Positioning: Frontier models for long-horizon agentic work and code

Moonshot AIKimi models, open-weight agentic performance at low cost60.0

Capability

Flagship model: Kimi K3 (kimi-k3)
Context window (tokens): 1,048,576tokens
Max output (tokens): ·
Image input: Yes
Audio in/out: No
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): –

Pricing

$/M input (/M tok): $3/M tok
$/M output (/M tok): $15/M tok
$/M cache read (/M tok): $0.3/M tok
Batch discount (%): –

Agentic

Tool use: Yes
Schema output: Partial
Effort control: Partial
Computer use: No
MCP support: No
Cache TTL: Automatic, no fixed TTL

Governance & continuity

Continuity policy: Dated model IDs (the -0905 style suffix) are used and remain addressable, which is better than DeepSeek's rolling aliases. No published deprecation policy or notice window. As with the other open-weight labs, the real guarantee is the licence, download the checkpoint and continuity is your problem, not theirs.
Notice period (days): ·
Pinnable versions: Yes
Zero retention: No
Trains on your data: Partial
AU region: No

Traction

API since: 2023

Positioning

Positioning: Open-weight agentic performance at open-weight prices

OpenRouterOne OpenAI-shaped key in front of hundreds of models59.6

Capability

Flagship model: Whichever upstream model you route to (400+ available)
Context window (tokens): –
Max output (tokens): –
Image input: Partial
Audio in/out: Partial
Open weights: Partial
OpenAI-compat API: Yes

Performance

TTFT p50: –
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): ·
Batch discount (%): 0%

Agentic

Tool use: Yes
Schema output: Partial
Effort control: Partial
Computer use: No
MCP support: No
Cache TTL: Passed through where the upstream supports caching

Governance & continuity

Continuity policy: Structurally the best hedge on this page and formally the weakest promise. OpenRouter guarantees nothing itself, a closed model vanishes the moment its lab retires it. But for open-weight models the same ID stays routable because several independent hosts serve it, and automatic failover means a single host going dark is invisible to your application. Buy it as insurance against host failure, not against model retirement.
Notice period (days): –
Pinnable versions: Partial
Zero retention: Partial
Trains on your data: No
AU region: No

Traction

API since: 2023

Positioning

Positioning: One key, one API shape, every model worth calling

Z.ai (GLM)GLM models with an unusually cheap flat-rate coding plan58.2

Capability

Flagship model: GLM-4.6
Context window (tokens): 200,000tokens
Max output (tokens): ·
Image input: Partial
Audio in/out: No
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $0.6/M tok
$/M output (/M tok): $2.2/M tok
$/M cache read (/M tok): ·
Batch discount (%): ·

Agentic

Tool use: Yes
Schema output: Partial
Effort control: Partial
Computer use: No
MCP support: No
Cache TTL: ·

Governance & continuity

Continuity policy: Versioned model names (GLM-4.5, 4.6) stay addressable after a successor ships, but there is no published deprecation policy or notice period. MIT-licensed weights are the fallback: every flagship generation so far has been released publicly, so a retired endpoint does not strand the model.
Notice period (days): ·
Pinnable versions: Yes
Zero retention: No
Trains on your data: Partial
AU region: No

Traction

API since: 2023

Positioning

Positioning: Coding-focused GLM models with a flat-rate subscription

GroqCustom LPU silicon serving open-weight models at extreme speed55.3

Capability

Flagship model: Curated open-weight models (Kimi K2, Llama 4, Qwen3, GPT-OSS)
Context window (tokens): –
Max output (tokens): –
Image input: Partial
Audio in/out: Partial
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: 250ms
Output tok/s (tok/s): 400tok/s

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): ·
Batch discount (%): –

Agentic

Tool use: Partial
Schema output: Partial
Effort control: No
Computer use: No
MCP support: No
Cache TTL: ·

Governance & continuity

Continuity policy: The weakest continuity of any host here. Models are added and removed from the console on a short cycle as Groq re-allocates LPU capacity, deprecations are announced with weeks rather than months of notice, and there is no dedicated-endpoint escape hatch for a model that gets pulled. Design for substitution: keep two model IDs configured and an eval you can rerun.
Notice period (days): –
Pinnable versions: Partial
Zero retention: Unknown
Trains on your data: No
AU region: No

Traction

API since: 2024

Positioning

Positioning: Custom LPU silicon for open-weight models at very low latency

CerebrasWafer-scale inference, the fastest tokens per second available55.2

Capability

Flagship model: Curated open-weight models (Qwen3, GLM, Llama, GPT-OSS)
Context window (tokens): –
Max output (tokens): –
Image input: No
Audio in/out: No
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: 200ms
Output tok/s (tok/s): 2,000tok/s

Pricing

$/M input (/M tok): –
$/M output (/M tok): ·
$/M cache read (/M tok): ·
Batch discount (%): ·

Agentic

Tool use: Partial
Schema output: Partial
Effort control: No
Computer use: No
MCP support: No
Cache TTL: ·

Governance & continuity

Continuity policy: Same fragility as Groq, from the same cause: a small curated menu sized to available wafer-scale capacity, with models added and removed as demand shifts. No published notice window. The models themselves are open-weight, so a removal costs you a re-benchmark and a host change rather than a rewrite.
Notice period (days): ·
Pinnable versions: Partial
Zero retention: Unknown
Trains on your data: No
AU region: No

Traction

API since: 2024

Positioning

Positioning: Wafer-scale inference, the highest tokens per second available

DeepSeekFrontier-adjacent models at a small fraction of Western prices53.3

Capability

Flagship model: DeepSeek-V4-Flash (deepseek-v4-flash)
Context window (tokens): 1,000,000tokens
Max output (tokens): 384,000tokens
Image input: No
Audio in/out: No
Open weights: Yes
OpenAI-compat API: Yes

Performance

TTFT p50: ·
Output tok/s (tok/s): ·

Pricing

$/M input (/M tok): $0.14/M tok
$/M output (/M tok): $0.28/M tok
$/M cache read (/M tok): $0.0028/M tok
Batch discount (%): 0%

Agentic

Tool use: Yes
Schema output: Partial
Effort control: Partial
Computer use: No
MCP support: No
Cache TTL: Automatic disk cache, no fixed TTL

Governance & continuity

Continuity policy: The worst on this page for a closed endpoint, and the best if you self-host. The API offers only rolling aliases, deepseek-chat silently points at whatever the current model is, so behaviour can change with no version to pin and no deprecation notice. The mitigation is the MIT licence: download the checkpoint and the version is yours permanently.
Notice period (days): –
Pinnable versions: No
Zero retention: No
Trains on your data: Yes
AU region: No

Traction

API since: 2023

Positioning

Positioning: Near-frontier quality at a fraction of the token price

yespartialnounknown

Sources shown beside each value · Learn how sourcing works

measured vendor-claimed community inferredExpand any row for the source, verification date, and caveat behind every cell.

Frontier LLM APIs compared, price, context, tool use and deprecation policy pricing, compared line by line

Where each one wins

Google Gemini

76.9

Workloads that genuinely need 500K+ tokens of context per request
Voice and video products wanting one model rather than a pipeline
Australian or regulated deployments that must pin inference to a region

The cheapest million-token window and only genuinely native audio, but the fastest-churning line-up here, so expect yearly migrations.

Amazon Bedrock

75.6

Regulated workloads that must stay inside an existing AWS perimeter
Australian data residency without self-hosting
Teams that need a documented model lifecycle to satisfy an auditor

Choose Bedrock for procurement and governance, not capability; accept lagging features and no MCP connector or automatic caching.

Mistral AI

75.2

EU data-residency requirements that rule out US-hosted inference
On-premise deployment with a commercial support contract
Teams that want an open-weight escape hatch under the hosted API

Rarely the capability leader, consistently the answer when jurisdiction is the blocker; check licences per model.

OpenAI

73.2

Multimodal products that need speech and image generation alongside text
Teams standardising on one vendor for everything rather than best-of-breed
Anyone who wants a dated snapshot they can actually pin

The widest surface area, with audio, images and text on one bill; pinnable snapshots but shorter observed notice.

Qwen

68.6

Finding an open-weight model at a specific size or modality
Fine-tuning where a permissive licence and a model-size ladder both matter
Asia-Pacific deployments already on Alibaba Cloud

The best open-weight range, a model at nearly every size and modality; the hosted Max tier is rarely the reason to be here.

xAI

64.5

Products that want real-time X/Twitter context in the model
Cheap evaluation of a frontier alternative with a one-line base-URL change
Consumer applications where governance requirements are light

Easy to try and strong on price-performance, but the thinnest governance here, so fine for consumer, hard for regulated work.

Meta Llama

64.3

Workloads that must survive any single vendor disappearing
On-premise or air-gapped deployments with a recognisable licence
Fine-tuning on your own data without a vendor in the loop

A model family, not a serious first-party endpoint; its value is continuity insurance, a model that cannot be deprecated.

Self-hosted (vLLM)

63.8

Data that legally cannot leave your infrastructure
Sustained high-utilisation inference where GPU-hours beat per-token billing
Guaranteeing a model version survives indefinitely

A reference point, not a recommendation; it wins on cost only at high steady load, so choose it for sovereignty or permanence.

Cohere

63.6

RAG pipelines where rerank quality drives the result
Deployments that must run inside your own VPC or data centre
Enterprises whose procurement needs a vendor agreement, not a credit card

Competes on retrieval, private deployment and contracts; its rerank and embeddings are the real draw, the chat API overpriced.

Together AI

62.9

Running DeepSeek or Kimi weights under a Western contract
Steady high-volume load that justifies a dedicated GPU endpoint
Fine-tuning open-weight models without owning hardware

The pragmatic way to run Chinese open weights on US infrastructure under a recognisable contract; dedicated endpoints shine at steady load.

Fireworks AI

61.9

Open-weight workloads where output must validate against a schema every time
Latency-sensitive serving without moving to custom silicon
Fine-tuned open models served on managed infrastructure

Hard to separate from Together, but stronger on constrained decoding and latency; pick it when output must parse every time.

Anthropic

61.0

Agent harnesses that call tools for minutes at a time
Code generation and review where first-try correctness pays for the token price
Teams that want cost tuned per route via graduated effort levels

The best platform for long tool loops, but continuity regressed to retirement dates and Fable 5 can refuse outright.

Moonshot AI

60.0

Agent loops on an open-weight budget
Self-hosting a tool-use-capable model under a permissive licence
Replacing a frontier model on the cheaper half of a mixed workload

Kimi K2 is the strongest case that open weights caught up on agentic work; reach it through Groq, Fireworks or Together.

OpenRouter

59.6

Keeping model choice a runtime decision rather than an architectural one
Benchmarking a dozen models without a dozen accounts
Failover when a single upstream host goes down

The highest-leverage integration: one endpoint, hundreds of models, failover; the trade is intersection features, not the union.

Z.ai (GLM)

58.2

Driving a coding agent all day on a predictable monthly cost
Cheap high-volume code completion and refactoring
Self-hosting an MIT-licensed coding model

The flat-rate coding plan is genuinely disruptive, turning all-day agent use into a fixed cost; quality trails only on hard reasoning.

Groq

55.3

Voice agents and any interface where perceived latency is the feature
High-volume cheap inference on well-known open-weight models
Fast transcription alongside text generation on one key

When instant responses are the product, Groq changes what you can build; not general-purpose, with a short, rotating menu.

Cerebras

55.2

Reasoning models where thinking-token latency is the bottleneck
Interactive tools that feel broken at 50 tokens per second
Benchmarking how much of your UX problem is actually latency

The fastest single-stream generation you can buy, by a qualitative margin; the catch is Groq's: short menu, no continuity guarantees.

DeepSeek

53.3

High-volume classification, extraction and summarisation
Cost-sensitive workloads where a 35x output-price gap against a volume-tier frontier model outweighs a small quality gap
Self-hosting a capable model under a genuinely permissive licence

The category's price floor with unmatched quality-per-dollar for bulk work; for regulated data, use the MIT weights via a Western host.

Which frontier LLM API should I build on?

Every provider quotes $/M tokens; almost none tell you how long the model lives
That continuity gap is the biggest hidden cost, so this page leads with it
Anthropic for agentic work: effort control, prompt caching and MCP are first-class
OpenAI for audio, image generation and text on one bill, or Responses API fluency
Google for cheap million-token context, native audio, or AU-resident Vertex deployment
Those three are not interchangeable; a tuned prompt lands differently on each
Chinese open-weight labs (DeepSeek, Moonshot, Qwen, Z.ai) reset the price floor
DeepSeek bills $0.42/M output vs $50 for Claude Fable 5, ~120x cheaper
Still ~35x cheaper than the volume tier most teams would actually use
For classification, extraction, summarisation and bulk rewriting, the quality gap is not worth the price gap
The catch is governance: permissive retention, no ZDR, rolling aliases that upgrade under you
For sensitive data, self-host the weights or route through a contractually clean host
OpenRouter is the cheapest insurance: one integration, hundreds of models, retired models often still reachable
Groq and Cerebras are latency instruments, not general-purpose substitutes
Worth an entire architecture for voice, autocomplete or interactive search; frustrating otherwise
Prices keep falling; context windows have stopped being the differentiator
The fight has moved to agentic fidelity: parallel tool calls, strict schema, durable caching
Assume today's model retires within two years
Build the abstraction layer now, keep a re-runnable eval set, treat every provider as replaceable

The pricing traps nobody puts on the pricing page

Headline $/M is the least useful number here; four things distort it
Output tokens dominate: a reasoning model burns tens of thousands of thinking tokens, billed at the output rate
A cheap-input, expensive-output provider is dear for reasoning, cheap for RAG answering
Model your actual input:output ratio before comparing
Cached input is where the money is: cache reads bill at ~10% of input at Anthropic and OpenAI
An agent loop resending its history is mostly cache reads
Caching is a prefix match; a timestamp, non-deterministic JSON, or a per-user tool list destroys it silently
Watch the cache-read token counter, not the invoice
Long-context surcharges: Google charges more above 200K tokens; others meter cache storage per hour
A million-token window at the short-context rate is not what you pay
Batch, but check the column first: where an async endpoint exists the discount is usually a flat 50%
Anthropic, OpenAI, Google, Mistral, Qwen, Bedrock and Together all publish that rate
DeepSeek offers no batch endpoint; its list price already sits below most batch rates
OpenRouter offers none either, a real cost if half your workload tolerates async
xAI, Moonshot, Z.ai, Cohere, Fireworks and Cerebras publish nothing confirmable
Groq documents a batch discount without a percentage we could stand behind
Where the halving exists and latency is negotiable, no negotiation beats it; where it is not, no committed spend conjures it

How to make a migration survivable

Assume an eighteen-month life for whatever you ship on
Providers that publish retirement dates help you; those that don't will break you
Pin a dated snapshot wherever offered, and record the pin in configuration, not code
This is getting harder: Anthropic's current line-up ships bare aliases, so you fall back on the retirement date
Keep the provider behind one seam: a per-provider module owning request construction, retries and parsing
Not an abstraction that pretends every model is the same; that fails on thinking blocks or tool-result shapes
Two implementations behind one interface is cheap insurance; a universal adapter is expensive fantasy
Keep a re-runnable eval set: 50-200 real inputs with graded expected outputs
Without it a forced migration is a multi-week vibe check; with it, a morning's work and a defensible number
Know your escape hatch first: for open weights it is total, vLLM serves them as long as you have GPUs
For closed models it is a router that may still have an upstream, or nothing at all

Deciding: three workloads, three answers

Bulk text processing (classification, extraction, tagging, translation, firehose summarising): price dominates, capability barely registers
DeepSeek, Qwen, Moonshot and Z.ai are the rational choices
Or Groq and Together for open weights with a Western contract
Add the batch endpoint if latency is negotiable
Never send sensitive data through a provider whose retention terms you haven't read
Agent harnesses (plan, call tools, read results, iterate for minutes): this is where frontier labs earn their price
You need well-formed tool calls at depth, enforced schema output, effort control, and caching that survives the loop
Anthropic and OpenAI are credible; Google is close and cheaper on long context
Budget for a long agentic turn running for minutes on one request
Interactive, latency-critical UI (voice, live search, inline completion): time to first token is the product
Cerebras and Groq are an order of magnitude ahead of first-party endpoints on tok/s
The constraint: you take whichever open-weight models they host
Design the feature around the model menu, not the reverse

What is frontier llm apis compared, price, context, tool use and deprecation policy?

Frontier LLM APIs compared, price, context, tool use and deprecation policy: On toolweight, Frontier LLM APIs compared, price, context, tool use and deprecation policy means the 18 tools benchmarked on this page, Anthropic, OpenAI, Google Gemini, xAI, Meta Llama, Mistral AI, DeepSeek, Qwen, Moonshot AI, Z.ai (GLM), Amazon Bedrock, Cohere, OpenRouter, Together AI, Fireworks AI, Groq, Cerebras, Self-hosted (vLLM), judged on the same 27 fields, from the same sources, on the same date. The question it exists to answer: Which frontier LLM API should I build on?

How does toolweight compare these?

Pricing here is stale within a fortnight; this category re-prices faster than any other on toolweight
Seven-day verification cadence, every price cell dated
Treat any cell dated more than a month ago as indicative; click through to the vendor first
Figures are pay-as-you-go list rates in USD per million tokens
Before batch, cache or committed-spend discounts, and before long-context surcharges
Where a vendor runs an unexpired introductory rate, the column carries list price and the note carries the intro rate and expiry
An introductory rate is what you actually pay today, so read the note before modelling a bill
Entries are providers, not models
Every price, context and capability figure describes that provider's current flagship, named in the Flagship model column
A cheaper flagship often sits beside a cheaper mid-tier that would serve you better
The killer column records notice before retirement, and whether you can pin a dated snapshot
Continuity is judged from published deprecation pages and observed retirements, not marketing
Where a provider publishes no policy, we say so rather than guess
Latency figures are community-reported; toolweight has not run its own TTFT harness, so most of that column is blank

Full methodology and sourcing policy →

Cite this comparisonCC-BY-4.0 · verified 2026-07-30

Anthropic, OpenAI and Google lead on capability, cheaper Chinese labs win on price, routers trade features for reach; pick on continuity, not headline price. — toolweight, https://toolweight.com/compare/llm-apis, verified 2026-07-30. Data from toolweight (https://toolweight.com), licensed CC-BY-4.0.

Frequently asked questions

What actually happens when a model I depend on is deprecated?

The alias stops resolving and requests 404
A pinned dated snapshot keeps serving until its published retirement date
A bare alias means you migrate today
Anthropic and OpenAI publish retirement dates months ahead; several providers publish nothing
Rolling aliases can change behaviour with no version bump

Is DeepSeek or Kimi really good enough to replace Claude or GPT?

For extraction, classification, summarisation, translation and bulk rewriting: yes
Roughly 1/35th the output cost of Anthropic's volume model, or 1/120th of its flagship
The honest bulk-work comparison is against the volume tier, not the flagship
For long agentic runs, sustained instruction following and first-try code, frontier labs stay ahead
Split the workload rather than picking one provider for everything

Does an OpenAI-compatible endpoint actually make switching easy?

It makes the transport identical and the semantics different
Chat completions port cleanly
Tool-call formats, effort parameters, cache accounting, thinking blocks and schema enforcement do not
Prompts tuned on one model regress on another
Budget a re-evaluation, not a config change

How much does prompt caching really save on an agent loop?

More than any other lever
Cache reads typically bill at 10-25% of the input rate
An agent loop resends the whole conversation each turn, so a long run is 80-90% cache reads
The trap: caching is a prefix match, so a timestamp or reordered tool list invalidates everything after it

Which providers can serve inference from an Australian region?

Amazon Bedrock (ap-southeast-2) and Google Vertex AI (australia-southeast1) serve frontier models in-country
First-party Anthropic, OpenAI and xAI endpoints route to US or EU by default
Residency controls vary by plan
Self-hosting is the only unambiguously AU-resident answer

Should I go through a router like OpenRouter instead of direct?

Use a router when model choice is a runtime decision, for outage fallback, or while benchmarking
Go direct when you depend on provider-specific features
That means cache breakpoints, computer use, batch endpoints, enterprise retention terms
Routers expose the intersection of upstream features, not the union

Why is the time-to-first-token column mostly empty?

Almost nobody publishes it
Self-reported latency is meaningless without a stated prompt length, region and concurrency
Figures shown are community measurements for speed-focused providers
toolweight will fill it only after running its own harness with published methodology