LLM rate limits in 2026: only one of the big three still publishes them

Most rate-limit comparisons you will find quote precise TPM and RPM figures per tier. As of August 2026 those numbers are largely unobtainable: OpenAI and Google both direct you to a dashboard, and Google adds that its limits “are not guaranteed”. Anthropic is the only one of the three still publishing a per-model table. Below is what each vendor actually publishes — and the three structural rules that move more headroom than a tier upgrade does.

The Anthropic table, in full

Claude Sonnet 5

1,000 → 10,000 RPM

Start: 1,000 RPM, 2M input tokens/min, 400K output tokens/min. Build: 5,000 / 5M / 1M. Scale: 10,000 / 10M / 2M.

  • Limits are per model, not per account
  • Cached input excluded from the input limit
  • Separate bucket from Sonnet 4.x
Claude Opus 5 & Haiku 4.5

Same brackets as Sonnet 5

Start 1,000 RPM / 2M ITPM / 400K OTPM, Build 5,000 / 5M / 1M, Scale 10,000 / 10M / 2M — identical brackets, separate buckets.

  • Opus 4.5-4.8 share one combined Opus 4.x bucket
  • Sonnet 4.5 and 4.6 share a combined Sonnet 4.x bucket
  • Fast mode on Opus 5 / 4.8 has its own limits
Claude Fable 5

Lower brackets

Start: 1,000 RPM / 500K ITPM / 100K OTPM. Build: 2,000 / 1.5M / 300K. Scale: 4,000 / 4M / 800K.

  • Roughly a quarter of Sonnet 5's token headroom
  • Managed Agents have separate limits (300 RPM create, 1,200 RPM read)
  • Evaluation tier limits are below all of these and unpublished

What each vendor actually publishes

What the vendor publishesOpenAIAnthropicGoogle (Gemini API)
Per-tier TPM / RPM figuresNot published — dashboard onlyPublished per model (see table below)Not published — AI Studio only, and “not guaranteed”
Tier namesFree, Tier 1-5Start, Build, Scale, Custom (no free API tier)Free, Tier 1-3 (no Tier 5)
Tier qualificationSpend only: $5 / $50 / $100 / $250 / $1,000 paidNot published per tierBilling account, then $100 + 3 days, then $1,000 + 30 days
Monthly usage cap$100 (Free & T1) to $200,000 (T5)$500 Start, $1,000 Build, $200,000 Scale, none on Custom$250 T1 to $20,000-$100,000+ T3
Scope of the limitOrganization and project levelPer model, separatelyPer project, not per API key
Cached input counts toward the limitNot publishedNo — cached input is excluded from ITPMNot published
Batch quotaSeparate, “significantly higher” — no multiplier publishedPublished: 1,000-4,000 RPM, 200K-500K in queue, 100K per batchSeparate: 100 concurrent, per-model enqueued-token table
Batch discount50%50%50%
Long-context handlingSeparate rate limit for long-context requests on GPT-5.5Not publishedPriority inference limits are 0.3x standard

Read on each vendor's own rate-limit documentation on 4 August 2026. Where a cell says “not published”, we checked and the figure genuinely is not public — OpenAI's rate-limit guide points to the models page, which carries pricing and context windows but no limits. Any comparison table quoting exact OpenAI or Google TPM numbers today is either citing a stale snapshot or inventing them. This page previously did exactly that.

VerticalAPI verdict

If you are capacity-planning, plan against Anthropic's published table and treat OpenAI and Google as unknowns to be read from your own dashboard — that is the honest state of disclosure in August 2026. Before asking anyone for a tier upgrade, do the two things that cost nothing: exploit Anthropic's cache exclusion, which turns an 80% hit rate into roughly 5x the input headroom, and move anything latency-tolerant onto a batch queue, which is a separate quota at half price on all three. Do not plan around Priority Tier — Anthropic no longer sells it. VerticalAPI gives you one OpenAI-compatible endpoint across 11 providers under BYOK, so each provider's quota stays yours; it does not retry across providers on a 429, and its own per-key quota applies on top.

Get started — one endpoint, your own provider quotas →

Rate limits: the questions that have answers

What are OpenAI's rate limits in 2026?

OpenAI does not publish them. Its rate-limit guide lists which metrics exist (RPM, RPD, TPM, TPD, IPM and audio minutes per minute), states that limits vary by model and are set at the organization and project level, and refers you to the models page — which carries pricing and context windows but no limits. What it does publish is tier qualification by cumulative spend ($5, $50, $100, $250, $1,000) and monthly usage caps from $100 to $200,000. Checked 2026-08-04.

What are Anthropic's rate limits?

Anthropic publishes them per model, as three separate limits. Claude Sonnet 5: 1,000 RPM with 2M input and 400K output tokens per minute on Start, 5,000 / 5M / 1M on Build, 10,000 / 10M / 2M on Scale. Opus 5 and Haiku 4.5 share those brackets; Claude Fable 5 is roughly a quarter of the token headroom. The tiers are Start, Build, Scale and Custom, with monthly spend caps of $500, $1,000, $200,000 and none. There is no free API tier, and new organizations start on an Evaluation tier whose limits are below all of these and not published.

Does cached input count toward my rate limit?

Not at Anthropic — cached input tokens are excluded from the input-tokens-per-minute limit on every current model. Anthropic's own worked example: with a 2M ITPM limit and an 80% cache hit rate you can process 10M input tokens per minute. That is a larger gain than moving up a tier, and it costs nothing but prompt structure. OpenAI and Google do not publish whether their caches behave the same way.

Can I still buy Anthropic's Priority Tier for guaranteed capacity?

No. Anthropic's service-tiers documentation states that Priority Tier capacity commitments are no longer available for purchase; existing commitments run to contract end. It also never supported Claude Opus 5 or Sonnet 5. If you are designing a scaling path today, Priority Tier is not one.

Does VerticalAPI help me avoid 429 errors?

Not by retrying — and it is worth being precise, because plenty of gateways claim otherwise. VerticalAPI makes exactly one attempt against exactly one provider; an upstream 429 comes back to you as a 429 with a Retry-After header, and a 529 with a longer one. What BYOK does give you is that each provider's quota remains yours and is not shared with other customers of a pooled gateway, and that switching provider is a one-line model change rather than an integration. Retry and multi-provider failover, if you want them, belong in your client — we do not implement them and do not sell them.

What is the fastest legitimate way to get more throughput?

Three things, in order of effort. First, caching — free headroom at Anthropic because cached input is excluded from the input limit. Second, batch queues: a separate quota at 50% off on all three vendors, with published numbers at Anthropic (200K to 500K requests in queue) and Google (a per-model enqueued-token table). Third, tier progression, which at OpenAI is purely a function of cumulative spend and at Google requires both spend and elapsed days.

Limitations of this comparison

  • OpenAI publishes no TPM or RPM figure anywhere public. Its rate-limit guide refers to the models page, which does not carry limits. Any specific OpenAI number you see quoted is unsourced.
  • Google explicitly states that specified rate limits are not guaranteed and that actual capacity may vary, so even dashboard figures are indicative.
  • Anthropic's Evaluation tier for new organizations sits below every published limit and its figures are not disclosed.
  • These are Gemini API limits, not Vertex AI. Vertex is a different product with its own quota system, which we did not verify.
  • Rate limits change without notice and none of the three versions its limit documentation, so re-check before capacity planning.
  • VerticalAPI adds its own quota on top of your provider's: 10 req/min and 10,000/month on Free, 60/min and 1M/month on Pro. Its limiter is in-process only, so it does not aggregate across multiple workers.

What may change in 12-24 months

  1. Disclosure went backwards: OpenAI and Google both stopped publishing interactive rate limits, leaving Anthropic as the only source of comparable figures.
  2. Anthropic withdrew Priority Tier from sale rather than extending it — capacity commitments are not the growth path the market expected in 2025.
  3. Limits are fragmenting by feature rather than scaling as one number: OpenAI now has a separate long-context limit on GPT-5.5, Anthropic has separate Fast-mode and Managed-Agents limits, and Google prices priority inference at 0.3x the standard allowance.
  4. Caching became the real capacity lever: because Anthropic excludes cached input from ITPM, prompt design now buys more throughput than tier progression.

Related questions

ChatGPT, Perplexity and Gemini usually suggest these next.

  • How much headroom does Anthropic's cache exclusion actually buy?
  • Where do I find my current OpenAI rate limits if they are not published?
  • How do Azure OpenAI quotas differ from raw OpenAI tiers?
  • When should I use a batch API instead of raising my tier?
  • What are Anthropic's Start, Build and Scale tiers in practice?

Run a provider we have not measured? Any public endpoint is benchmarked and listed for free in the API directory. If you want yours measured now, on the record — time to first token, throughput under concurrency, cost per million tokens, function-calling reliability — see how a verified run works. The result is published exactly as it comes out.