API model string
claude-opus-4-8Context window
1M tokens (no surcharge)
Max output 128K tokens (sync), 300K tokens (batch)
- Knowledge cutoff
- not published
- Released
- May 28, 2026
- Modalities
- text + image in, text out
Last verified July 10, 2026
Rate limits by tier
Anthropic replaced its numeric Tier 1–4 system with Start, Build, Scale (+ Custom) on June 26, 2026. Monthly spend caps: Start $500, Build $1,000, Scale $200,000. As of June 2026, Opus 4.x now shares the same RPM/ITPM/OTPM ceilings as Sonnet 4.x/5 at every tier (previously Opus had higher ceilings). Opus 4.x models share one org-level pool. Figures are per Anthropic docs via third-party trackers — verify actual limits in your console at platform.claude.com/settings/limits.
| Tier | Requirements | RPM | TPM | RPD | Concurrent | Notes |
|---|---|---|---|---|---|---|
| Start | First API key or cumulative deposit under $40 | ~50 (per third-party trackers — confirm in your console/dashboard) | ~30,000 ITPM (legacy Tier-1 Opus reference; anecdotal post-Colossus reports cite ~500,000 ITPM — per third-party trackers, confirm in your console/dashboard) | not published | not published | Entry tier; adequate for testing and prototyping. Acceleration limits can trigger 429 on sudden spikes even within the ceiling. |
| Build | Cumulative spend above ~$40; automatic promotion | not published | not published | not published | not published | Automatic on crossing spend threshold. RPM and ITPM ceilings scale above Start. Spend cap $1,000/month. |
| Scale | Cumulative deposit threshold (~$200 equivalent); spend cap $200,000/month | ~4,000 (per third-party trackers — confirm in your console/dashboard) | ~2,000,000 ITPM; ~400,000–800,000 OTPM (per third-party trackers — confirm in your console/dashboard) | not published | not published | Primary production planning ceiling for Opus 4.8 workloads. Cache reads do NOT count toward ITPM for Opus 4.8. Higher limits available via Custom tier through Anthropic sales. |
| Custom | Contact Anthropic sales | not published | not published | not published | not published | No spend cap. Negotiated RPM/ITPM/OTPM above Scale ceilings. Dedicated support SLA. Fast Mode has separate limits tracked via anthropic-fast-* headers. |
Swipe the table sideways to see every limit column.
- 1.Token-bucket replenishment is continuous. Acceleration limits can trigger 429 on sudden spikes even within tier ceilings — ramp traffic gradually over minutes, not seconds.
- 2.Cache reads do NOT count toward ITPM for Opus 4.8 (current model). This is the primary throughput optimization for large-context Opus workloads.
- 3.Opus 4.x org pool is shared across all Opus 4.x model versions at the org level. Fast Mode ($10/$50/MTok) has its own separate pool tracked via anthropic-fast-* headers.
Limits verified against the Anthropic docs, July 10, 2026.
Token pricing
What you pay per million tokens (USD). Input and output are billed separately.
Input
$5.00
per 1M tokens
Output
$25.00
per 1M tokens
Cached input
$0.50
per 1M cached
Batch discount
−50%
on batch jobs
- Cache read = $0.50/MTok (0.1× base input $5.00). 5-minute cache write = $6.25/MTok (1.25× base). 1-hour cache write = $10.00/MTok (2× base). Cache reads excluded from ITPM.
- Batch API (50% off): $2.50 input / $12.50 output per MTok. Async within 24 hours. Max 300K output tokens per request with output-300k-2026-03-24 beta header.
- Fast Mode: $10.00 input / $50.00 output per MTok (2× standard). Reserve for latency-critical user-facing paths. Fast Mode limits are tracked separately via anthropic-fast-* headers.
- Extended thinking tokens billed as output (can multiply output cost 3–5×). Disable for non-reasoning tasks.
- US-only data residency (inference_geo: 'us') adds 1.1× surcharge on Opus 4.6 and later. Omit unless legally required.
- Web search tool adds $10 per 1,000 searches on top of token costs. Cache tool-calling contexts to minimize repeated searches.
- No long-context surcharge on 1M-context Opus 4.8. No free API tier. No first-party embeddings or fine-tuning.
Side-project with Opus reasoning
$50
per month
Assumptions
5M input tokens / 1M output tokens per month, no caching
5M × $5.00/MTok = $25 input + 1M × $25.00/MTok = $25 output = $50/mo
Mid-scale app with caching
$411
per month
Assumptions
60M input / 12M output per month, 70% of input served from cache
18M × $5.00/MTok = $90 (non-cached) + 42M × $0.50/MTok = $21 (cache reads) + 12M × $25.00/MTok = $300 output = $411/mo
High-volume production
$4,000
per month
Assumptions
400M input / 80M output per month, no caching
400M × $5.00/MTok = $2,000 input + 80M × $25.00/MTok = $2,000 output = $4,000/mo
Run your own numbers
Drag your real monthly token volumes and watch the bill update live — priced against rival models at the same usage.
Estimated claude-opus-4-8 spend
$275/mo
Input: $150
Output: $125
30M in × $5.00 + 5M out × $25.00 = $275
Same volume, priced across models
- Grok 4Cheapest$90
- Claude 4 Sonnet$110
- claude-opus-4-8This model$275
- GPT-5$300
Rivals priced at their published input/output rates for the same monthly volumes. Prompt caching is model-specific, so it is applied to claude-opus-4-8 only. Estimates for comparison; real bills vary with request shape and long-context surcharges.
claude-opus-4-8 vs the alternatives
Claude Opus 4.8 competes with Claude Sonnet 5 (lower-cost alternative), GPT-5.5 (OpenAI flagship), and Grok 4.5 (xAI) on context, pricing, and reasoning capability.
| Aspect | claude-opus-4-8 | Claude 4 Sonnet | GPT-5 | Grok 4 |
|---|---|---|---|---|
| Context window | 1M tokens | 1M tokens | ~1M tokens (922K input) | 500K tokens |
| Input $/MTok | $5.00 | $2.00 | $5.00 | $2.00 |
| Output $/MTok | $25.00 | $10.00 | $30.00 | $6.00 |
| Cached input $/MTok | $0.50 | $0.20 | $0.50 | $0.50 |
| Batch discount | 50% | 50% | 50% | not available at launch |
| Max output (sync) | 128K | 128K | 128K | not published |
| Long-context surcharge | none | none | 2× above 272K | not published above 200K |
| Coding/reasoning benchmark | Current premium flagship (Anthropic claim — per third-party trackers — confirm in your console/dashboard) | Default price/performance | GPT-5 family | Grok 4.5 #4 on Intelligence Index |
| Throughput ceiling (OTPM at Scale) | ~400K–800K OTPM | ~400K–800K OTPM | ~30M OTPM (OpenAI Tier 5) | not published |
Swipe the table sideways to see every model.
Hitting a 429? The playbook
The exact errors you'll see
rate_limit_errortype: "rate_limit_error"HTTP 429 Too Many RequestsWhy it happens & how to fix it
RPM burst: sending multiple Opus requests in rapid succession
Spread requests evenly. Anthropic's token-bucket uses acceleration limits that fire on rapid bursts even within tier ceilings. Ramp linearly over minutes, not seconds — especially important for Opus given its higher token costs per request.
ITPM saturation from large 1M-context prompts
Activate prompt caching. Cache reads do NOT count toward ITPM for Opus 4.8. Use cache_control breakpoints on repeated system prompts and long contexts. Use 1-hr cache writes ($10.00/MTok) for persistent contexts across sessions.
OTPM exhaustion from long Opus outputs
Reduce max_tokens per request. Route non-latency-sensitive jobs to Batch API ($2.50/$12.50 per MTok, async within 24h, up to 300K output tokens per request). This removes OTPM pressure from the real-time path.
Fast Mode limits hit separately from standard limits
Fast Mode ($10/$50 per MTok) has its own separate rate limit pool tracked via anthropic-fast-* headers. Parse anthropic-fast-requests-remaining and anthropic-fast-tokens-remaining when Fast Mode is active. Reserve Fast Mode for SLA-bound user-facing paths only.
Org Opus pool exhausted by multiple teams
Check the Rate Limits API (GET /v1/rate-limits, launched April 25, 2026) to see remaining headroom per dimension. Partition API keys across teams or time-box heavy workloads to off-peak hours.
Retry strategy
On receiving a 429, read the retry-after header (seconds to wait). Also parse anthropic-ratelimit-requests-remaining, anthropic-ratelimit-tokens-remaining, and anthropic-ratelimit-reset-tokens to determine whether you are RPM-bound or TPM-bound. Use exponential backoff with jitter: start at 1 second, double on each retry, cap at 64 seconds. Do not retry immediately on 429.
1const ANTHROPIC_API_URL = "https://api.anthropic.com/v1/messages";23async function callOpusWithRetry(4 payload: object,5 maxRetries = 56): Promise<Response> {7 let attempt = 0;8 let delayMs = 1000;910 while (attempt <= maxRetries) {11 const res = await fetch(ANTHROPIC_API_URL, {12 method: "POST",13 headers: {14 "x-api-key": process.env.ANTHROPIC_API_KEY!,15 "anthropic-version": "2023-06-01",16 "content-type": "application/json",17 },18 body: JSON.stringify(payload),19 });2021 if (res.status !== 429) return res;2223 const retryAfter = res.headers.get("retry-after");24 const reqsRemaining = res.headers.get("anthropic-ratelimit-requests-remaining");25 const toksRemaining = res.headers.get("anthropic-ratelimit-tokens-remaining");26 const resetToks = res.headers.get("anthropic-ratelimit-reset-tokens");2728 console.warn(29 `429 on attempt ${attempt + 1}/${maxRetries + 1}. ` +30 `Requests left: ${reqsRemaining}, Tokens left: ${toksRemaining}, ` +31 `Token reset: ${resetToks}, Retry-After: ${retryAfter}s`32 );3334 const waitMs = retryAfter35 ? parseInt(retryAfter, 10) * 100036 : delayMs + Math.random() * 500;3738 await new Promise((r) => setTimeout(r, waitMs));39 delayMs = Math.min(delayMs * 2, 64000);40 attempt++;41 }4243 throw new Error(`Exceeded max retries (${maxRetries}) on 429`);44}How to raise your limits
The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.
Start
ImmediateAutomatic on creating your first API key or depositing under $40 cumulative. No application required.
Unlocks: ~50 RPM (per third-party trackers). Spend cap $500/month. Adequate for testing Opus capabilities but not production workloads.
Build
Automatic within hours to days of crossing thresholdAutomatic when cumulative spend crosses ~$40. Monitor at platform.claude.com/settings/limits.
Unlocks: RPM and ITPM ceilings scale above Start. Spend cap $1,000/month. Suitable for early production with limited concurrent users.
Scale
Automatic within days of crossing thresholdAutomatic when cumulative deposit crosses ~$200 equivalent. Check platform.claude.com/settings/limits for your current position or query programmatically via GET /v1/rate-limits.
Unlocks: ~4,000 RPM / ~2,000,000 ITPM / ~400,000–800,000 OTPM (per third-party trackers — confirm in console). Spend cap $200,000/month. This is the primary production ceiling.
Custom
Days to weeks to negotiateContact Anthropic sales. Discuss volume commitments, enterprise agreements, and BAA if required. Annual volume commitments can accelerate Custom access.
Unlocks: No spend cap. Negotiated RPM/ITPM/OTPM above Scale. Dedicated support SLA. Fast Mode quota can be separately negotiated.
Cut your token spend
Prompt caching
−90% cost on repeated context; cache reads excluded from ITPM — dual throughput benefitCache reads = $0.50/MTok (0.1× base $5.00) AND excluded from ITPM. Use cache_control breakpoints on repeated system prompts and conversation history. Use 1-hr cache writes ($10.00/MTok) for long-running sessions where the same context appears repeatedly.
Batch API for non-realtime jobs
50% off ($2.50/$12.50 per MTok); removes OTPM pressure from real-time pathRoute all non-latency-sensitive work — evaluation runs, bulk extraction, nightly jobs — to Batch API. Async within 24h. Supports up to 300K output tokens per request with output-300k-2026-03-24 beta header.
Ramp traffic gradually
Avoids acceleration-triggered 429s that fire within tier ceilingsAnthropic token-bucket uses acceleration limits separate from tier ceilings. Smooth request curves over minutes. Never burst from 0 to max rate — this is especially critical for Opus given higher per-request cost.
Reserve Fast Mode for user-facing paths only
Avoids 2× ($10/$50) cost on non-latency-critical workFast Mode doubles standard Opus pricing to $10/$50 per MTok. Reserve for SLA-bound user-facing requests; route batch and background jobs to standard mode or Batch API.
Disable extended thinking on non-reasoning tasks
Avoids 3–5× output cost multiplierExtended thinking tokens bill as output tokens. Enable only when deep multi-step reasoning is genuinely required. For summarization, extraction, or classification tasks, disable thinking.
Skip US-only data residency unless required
Saves 10% on applicable Opus versionsinference_geo: 'us' adds 1.1× surcharge on Opus 4.6 and later. Omit unless you have a specific legal or compliance requirement.
Use Batch API for 300K-output tasks
Avoids stream timeouts and reduces per-request cost 50%Opus 4.8 max_output is 128K tokens in sync mode. For tasks requiring up to 300K output tokens, use Batch API with the output-300k-2026-03-24 beta header — both larger outputs and lower cost.
Cache web search tool results
Reduces $10/1K searches surcharge on repeated queriesThe web search tool charges $10 per 1,000 searches in addition to token costs. Cache tool-calling contexts with search results so the same search is not repeated across sessions or users.
Frequently asked questions
Is claude-opus-4 still working?
No. The claude-opus-4 API string stopped accepting requests on June 15, 2026. Any calls using that model string will fail. You must migrate to claude-opus-4-8, which is Anthropic's current premium flagship and the GA successor since May 28, 2026.
How do I increase Claude Opus 4.8 rate limits?
Rate limits increase automatically: ~$40 cumulative spend moves you to Build, ~$200 cumulative unlocks Scale (~4,000 RPM / ~2M ITPM per third-party trackers). Above Scale, contact Anthropic sales for Custom tier with negotiated limits and no spend cap. Monitor at platform.claude.com/settings/limits or query GET /v1/rate-limits programmatically.
What is the difference between Claude 4 Opus and Claude Opus 4.8?
Claude 4 Opus (claude-opus-4) was an earlier snapshot that stopped serving June 15, 2026. Claude Opus 4.8 (claude-opus-4-8) is the current GA model in the same Opus 4.x family, released May 28, 2026. They are in the same org-level pool and Opus 4.x limits apply to both. Use claude-opus-4-8 for all new work.
Is the Claude Opus 4.8 API free?
No. There is no free tier for Claude Opus 4.8. All token usage is billed from the Start tier onward. Standard pricing is $5.00 input / $25.00 output per MTok. Batch API offers 50% off ($2.50/$12.50), and cache reads are $0.50/MTok.
How does Claude Opus 4.8 compare to GPT-5 pricing?
Claude Opus 4.8 is $5.00/$25.00 per MTok vs GPT-5.5 at approximately $5.00/$30.00. Opus 4.8 has a meaningful output cost advantage ($25 vs $30 per MTok) and no long-context surcharge on 1M context, while GPT-5.5 charges 2× above 272K tokens. GPT-5.5 offers a much higher OTPM ceiling (~30M at Tier 5 vs ~800K for Opus at Scale), making it better suited for throughput-critical workloads.
What causes Claude API 429 errors on Opus?
The 429 rate_limit_error on Opus 4.8 means you exceeded RPM, ITPM, or OTPM limits. Large 1M-context prompts are the most common ITPM trigger. The fix: activate prompt caching (cache reads excluded from ITPM), reduce max_tokens, or route batch work to Batch API. Check anthropic-ratelimit-tokens-remaining and anthropic-ratelimit-requests-remaining headers to identify which limit you hit.
Does Claude Opus 4.8 support Priority Tier for lower latency?
Yes. Unlike Claude Sonnet 5 (which is NOT on Priority Tier), Claude Opus 4.8 supports Priority Tier access. Check with Anthropic sales when negotiating a Custom tier agreement if low-latency priority routing is a requirement for your use case.
How can I reduce Claude Opus API costs for high-volume workloads?
The top levers: (1) Prompt caching — cache reads cost $0.50/MTok vs $5.00 standard and are excluded from ITPM. (2) Batch API — 50% off at $2.50/$12.50 for async jobs. (3) Model routing — use Haiku 4.5 ($1/$5) or Sonnet 5 ($2/$10) for classification/routing and escalate to Opus 4.8 only for tasks requiring deep reasoning. RapidDev's engineers specialize in multi-tier Anthropic model routing strategies — reach out at rapidevelopers.com/contact for a free scoping call.
What is Fast Mode for Claude Opus 4.8?
Fast Mode is a high-speed inference option for Claude Opus 4.8 priced at $10.00 input / $50.00 output per MTok — 2× standard price. It has its own separate rate limit pool tracked via anthropic-fast-* headers. Reserve Fast Mode only for latency-critical user-facing paths where response time directly impacts UX; route all background and batch work through standard mode.
We build AI apps that don't hit rate limits
- Retry, backoff & caching built in
- Multi-provider fallback routing
- Fixed price, you own the code
30-min call. No commitment.