How to Use Claude Opus 5: Benchmarks, Effort, and Cost
How to use Claude Opus 5: the full official benchmark table, the effort ladder that decides your bill, two breaking API changes, and the migration steps.
Anthropic launched Claude Opus 5 on July 24, 2026, describing it as a step-change improvement over Claude Opus 4.8[6]. It lists at exactly the same price as the model it succeeds: $5 per million input tokens and $25 per million output[3]. So the question of how to use Claude Opus 5 well has almost nothing to do with the list price, which did not move, and almost everything to do with a dial most teams never touch.
That dial is effort. Instead of publishing one number per benchmark, Anthropic published score-versus-cost curves across all five effort levels[1]. The spread between the cheapest and the most expensive rung on those curves is wider than anything a prompt rewrite will recover, and the official guidance is to sweep the ladder against your own evals rather than turn everything up[2].
TL;DR
- Same list price, different model. $5 / $25 per million tokens, unchanged from Opus 4.8[2]. On reAPI the same model runs $2.40 / $12.00, which is 48% of Anthropic's published rate.
- Thinking is on by default now. On Opus 4.8, omitting the
thinkingfield meant no thinking. On Opus 5 the same request thinks[2]. Sincemax_tokenscaps thinking plus response text together, tight budgets carried over will truncate. - Disabling thinking is capped at effort
high. Sendingthinking: {"type": "disabled"}withxhighormaxreturns a 400, checked per request[2]. - Opus 5 does not win everything. Anthropic's own table has it losing agentic coding on DeepSWE v1.1 to GPT-5.6 Sol, 68.8% against 72.7%[1].
- Do not inherit effort settings. Anthropic says
lowandmediumnow produce strong quality at a fraction of the tokens and latency, and tells you to start at the defaulthighand move in both directions against your own evals[2]. - Priority Tier does not cover it, and it sits in its own rate-limit bucket rather than the shared Opus pool[5][4].
What Claude Opus 5 is
Anthropic positions Claude Opus 5 for complex agentic coding and enterprise work[2]. The API identifier is claude-opus-5, with no date suffix to append.
| Property | Claude Opus 5 |
|---|---|
| API model ID | claude-opus-5 |
| Context window | 1M tokens, both the default and the maximum |
| Max output | 128k tokens |
| List price | $5 / $25 per million input / output tokens |
| Fast mode | $10 / $50, Claude API only |
| Thinking | On by default, adaptive |
| Effort levels | low, medium, high, xhigh, max, default high |
| Prompt cache minimum | 512 tokens, down from 1,024 |
There is no smaller context variant[2]. The 1M window is what you get. Beyond the Claude API, the model is available on Amazon Bedrock as anthropic.claude-opus-5, on Google Cloud as claude-opus-5, and on Microsoft Foundry[2]. Fast mode is the exception: it runs on the Claude API only, not on any of the three cloud platforms[2].
The full official benchmark table
This is Anthropic's published comparison, transcribed in full, including the rows Opus 5 does not win[1]. Where the Fable 5 column carries a different model name, that is Anthropic's own annotation.
| Benchmark | Opus 5 | Fable 5 | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|
| Agentic terminal coding (Frontier-Bench v0.1) | 43.3% | 33.7% | 21.1% | 34.4% |
| Knowledge work (GDPval-AA v2, Elo) | 1861 | 1747 | 1593 | 1736 |
| Novel problem-solving (ARC-AGI-3) | 30.2% | — | 1.5% | 7.8% |
| Agentic search (BrowseComp) | 90.8% | 87.4% | 84.3% | 90.4% |
| Multidisciplinary reasoning (HLE, no tools) | 56.3% | 56.5% | 49.8% | — |
| Multidisciplinary reasoning (HLE, with tools) | 64.7% | 63.9% | 57.9% | — |
| Computer use (OSWorld 2.0) | 70.6% | 66.1% | 55.7% | 62.6% |
| Agentic coding (DeepSWE v1.1) | 68.8% | 69.7% | 59.0% | 72.7% |
| Agentic coding (FrontierCode v1.1, Main) | 53.4% | 53.5% | 46.5% | 47.5% |
| Business workflows (AutomationBench) | 26.0% | 17.4% | 17.0% | 18.1% |
| Legal (Legal Agent Benchmark, held-out) | 11.7% | 13.3% | 10.4% | 2.5% |
| Health (HealthBench Professional) | 59.8% | 66.0% (Mythos 5) | 57.4% | 60.5% |
| Biology (BioMysteryBench, hard) | 49.4% | 46.5% | 42.4% | — |
| Biology (BioMysteryBench, human-solved) | 90.1% | 89.0% (Mythos 5) | 88.5% | — |
One methodology note before anyone quotes these numbers. On the companion chart that breaks Frontier-Bench down by effort level, Anthropic states that those results come from an internal run on the mini-SWE-agent harness with a GKE backend, averaging reward over five attempts per task, and that Opus 4.8 served as the fallback when safety classifiers refused a request from Opus 5 or Fable 5[1]. These are vendor-run evaluations. Read them as a directional signal, not a neutral audit.
Where it wins and where it loses
Opus 5 takes nine of the fourteen rows. The margins on the wins are not subtle. Frontier-Bench goes from 21.1% on Opus 4.8 to 43.3%, slightly more than double. ARC-AGI-3 goes from 1.5% to 30.2%, and the nearest competitor in that row is GPT-5.6 Sol at 7.8%. AutomationBench sits at 26.0% against a cluster of 17% to 18%[1].
It loses the other five, and those are more useful for picking a model than the wins are.
- DeepSWE v1.1, agentic coding: 68.8% against GPT-5.6 Sol's 72.7%. This is the clearest loss on the board, and it lands on the workload Opus 5 is marketed for. Anthropic left the row in.
- HealthBench Professional: 59.8%, behind both Mythos 5 at 66.0% and GPT-5.6 Sol at 60.5%.
- Legal Agent Benchmark: 11.7% against Fable 5's 13.3%. Both numbers are low enough that the benchmark is mostly measuring distance from the goal.
- HLE without tools (56.3% against 56.5%) and FrontierCode Main (53.4% against 53.5%) are ties inside the noise, but they do rule out the claim that Opus 5 dominates Fable 5 across the board.
The pattern: Opus 5 wins where a task runs long, uses tools, and requires holding state across many steps. It ties or loses on single-shot hard questions.
The effort ladder decides your bill
Anthropic publishes the full score-versus-cost ladder rather than one number per benchmark, and reading it changes how you configure the model[1].
Three things follow from it.
The top rung is not the recommended starting point. Anthropic's instruction is to start at the default, high, and adjust in either direction based on your evals: step down where quality holds to save tokens and latency, step up for the most demanding work[2]. Note where that sentence starts. The recommended entry point is the middle of the ladder, not the top of it, and effort now carries more weight because Opus 5 converts extra effort into better results more reliably than any earlier Opus model[2].
Low and medium are stronger than they used to be. Anthropic explicitly calls out efficiency at lower effort, saying low and medium produce strong quality at a fraction of the tokens and latency of higher settings[2]. Effort defaults carried over from Opus 4.8 are the wrong starting point.
Effort is the cost lever, not the prompt. The official guidance is to start at the default high, then move in either direction against your own evals: step down where quality holds to save tokens and latency, step up for the most demanding work[2]. At xhigh or max, set a large max_tokens so the model has room to think and act across subagents and tool calls[2].
Two breaking changes under the hood
The request surface is mostly inherited from Opus 4.8. Sampling parameters are still rejected, fixed thinking budgets are still rejected, and last-assistant-turn prefill still fails. Two things did change, and both break code carried over unmodified.
Thinking runs when you omit the parameter. On Opus 4.8, a request without a thinking field ran without thinking. On Opus 5, the same request thinks, with the model deciding when and how much on each turn[2]. The wire value did not change; thinking: {"type": "adaptive"} is still valid and equivalent to the default. What changed is what happens when you say nothing. Because max_tokens is a hard limit on thinking plus response text together, Anthropic tells you to revisit it for any workload that ran without thinking before[2]. A budget tuned for visible output alone will now truncate mid-answer.
Turning thinking off requires effort high or below. thinking: {"type": "disabled"} paired with xhigh or max returns a 400. The check runs on every request, so a session that succeeded at high starts failing the moment a later call raises effort while thinking is still disabled[2].
Anthropic also documents what goes wrong with thinking off: the model can write a tool call into its text output instead of emitting a tool_use block, or leak internal XML tags into the visible response[2]. The first failure mode is the dangerous one in an agent loop, because the turn ends cleanly and the call never executes. The official recommendation is to keep thinking on and control cost with a lower effort level instead.
Three additions are worth knowing about:
- Mid-conversation tool changes (beta header
mid-conversation-tool-changes-2026-07-01) let you add or remove tools between turns while preserving the prompt cache, instead of pinning a tool list for the life of a session[2]. - A
"default"fallbacks mode (beta headerserver-side-fallback-2026-07-01) applies Anthropic's recommended fallback models by refusal category rather than a list you maintain[2]. - The prompt cache minimum dropped to 512 tokens from 1,024, so prompts previously too short to cache now create cache entries with no code changes[2].
What actually moves the bill
Here is the complete official rate card[3]:
| Item | Price per MTok |
|---|---|
| Input | $5 |
| Output | $25 |
| Cache hits and refreshes | $0.50 |
| Cache write (5-minute) | $6.25 |
| Cache write (1-hour) | $10 |
| Batch input | $2.50 |
| Batch output | $12.50 |
| Fast mode input | $10 |
| Fast mode output | $50 |
The Batch API runs at half the standard rate on both dimensions, and fast mode doubles both[3]. Anthropic describes Opus 5 as delivering frontier intelligence at half the cost of Claude Fable 5[2], which checks out against the rate card: Fable 5 lists at $10 / $50[3].
Three levers matter more than the per-token rate.
Effort. The published cost-per-attempt spread across the ladder is wide enough that moving one route from xhigh down to medium saves more than any prompt rewrite[1].
Verbosity. Anthropic states that default responses and written deliverables run longer on Opus 5[2]. Thinking tokens bill as output, and so does the extra prose. Lowering effort reduces thinking but does not reliably shorten visible output, so write an explicit brevity instruction.
Verification scaffolding you can now delete. Opus 5 verifies its own work without being told to, and Anthropic says instructions carried over from earlier models, like "include a final verification step" or "use a subagent to verify", cause over-verification[2]. Deleting them lowers token spend with no quality tradeoff, which is a rare kind of optimization.
Rate limits and Priority Tier
Two operational facts that catch teams during rollout.
Priority Tier does not cover Claude Opus 5. Anthropic's service-tier documentation states that Priority Tier is supported on all available Claude models except Claude Mythos 5, Claude Mythos Preview, Claude Opus 5, and Claude Sonnet 5[5]. If your capacity planning assumed Priority Tier coverage, it does not apply here.
Opus 5 also has its own rate-limit bucket. The combined Opus limit applies to traffic across Opus 4.8, 4.7, 4.6, and 4.5; Opus 5 is not part of that pool[4]. That is good news during a migration, since Opus 5 traffic does not eat the budget your existing Opus 4.8 routes depend on. Fast mode has dedicated limits again separate from standard Opus limits, and returns a 429 with a retry-after header when exceeded[4].
How to use Claude Opus 5: a migration checklist
Anthropic's migration guidance reduces to a model string change plus a review of the two behavior changes[2]. In practice the list is a little longer.
- Change the model string to
claude-opus-5. - Audit every route that disables thinking. Either keep thinking disabled and drop effort to
highor below, or keep the effort level and remove thethinkingfield. - Raise
max_tokenson routes that never set athinkingfield. They think now, and the budget covers both. - Re-run an effort sweep against your own evals. Start at the default
highand move in both directions rather than inheriting Opus 4.8 settings. - Delete verification instructions from prompts and verification steps from harnesses.
- Add an explicit brevity instruction to user-facing routes.
- Recheck short prompts you had written off as uncacheable against the new 512-token minimum.
- Handle refusals as responses, not errors. Safety classifiers can decline a request and return a stop reason rather than an HTTP error, so code that reads the first content block unconditionally will break.
Calling Claude Opus 5 alongside your other models
Everything above points at the same operational problem. This model has five effort levels, its own rate-limit bucket, no Priority Tier, and a benchmark table where it loses agentic coding to GPT-5.6 Sol and health to Mythos 5. The correct configuration is not one setting. It is a routing decision per workload, and it changes every time a vendor ships.
That is the layer worth building deliberately. A code-review pass and a bulk extraction job have no business running at the same effort, and a route that Opus 5 loses on has no business staying on Opus 5 out of habit. Handling that with one SDK client per vendor, one key per vendor, and one billing surface per vendor is where the cost of multi-model work actually accumulates.
reAPI exposes Claude Opus 5 through an OpenAI-compatible /v1/chat/completions endpoint, so the client you already use for GPT and Gemini calls it too:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_REAPI_KEY",
base_url="https://api.reapi.ai/v1",
)
resp = client.chat.completions.create(
model="claude-opus-5",
messages=[{"role": "user", "content": "Refactor this module and add tests."}],
max_tokens=16000,
stream=True,
)Rates on the gateway are $2.40 per million input tokens and $12.00 per million output, which is 48% of Anthropic's published $5 / $25. The full request surface, including output_config.effort and the thinking rules above, is documented on the reapi.ai/docs/claude-opus-5 page, and current rates live on reapi.ai/models/claude-opus-5.
None of that removes the routing decision. It removes the integration tax on making one.
FAQ
Is Claude Opus 5 better than Claude Fable 5?
On nine of the fourteen rows in Anthropic's published table, yes, and at half Fable 5's list price[3]. But Fable 5 still takes HLE without tools, FrontierCode Main, and the Legal Agent Benchmark, and the HealthBench Professional row goes to Mythos 5[1].
Does Claude Opus 5 cost more than Opus 4.8?
No. Both list at $5 per million input tokens and $25 per million output[3]. Anthropic shipped the capability jump without a price change[2].
What is the biggest breaking change from Opus 4.8?
Thinking is on by default. A request that omitted the thinking field ran without thinking on Opus 4.8 and thinks on Opus 5, which changes both cost and max_tokens behavior. Second is that thinking: {"type": "disabled"} now returns a 400 at effort xhigh or max[2].
Which effort level should I use?
Start at the default, high, then sweep in both directions against your own evals[2]. Anthropic frames max as the top tier for the deepest possible reasoning rather than a better default, and separately notes that low and medium hold up at a fraction of the tokens and latency on this model[2].
What is the Claude Opus 5 context window?
1M tokens, which is both the default and the maximum. There is no smaller context variant. Maximum output is 128k tokens[2].
Is Claude Opus 5 covered by Priority Tier?
No. Priority Tier covers all available Claude models except Mythos 5, Mythos Preview, Opus 5, and Sonnet 5[5]. Opus 5 also sits in its own rate-limit bucket rather than the combined Opus 4.x pool[4].
Can I still turn thinking off?
Yes, at effort high or below. Anthropic advises against it, documenting that the model may write tool calls into visible text or leak internal tags when thinking is disabled, and recommends lowering effort instead[2].
How do I call Claude Opus 5 with the OpenAI SDK?
Point the OpenAI client at https://api.reapi.ai/v1, set model to claude-opus-5, and send a normal chat request. The endpoint is OpenAI-compatible, so no SDK rewrite is needed.
Picking an effort level and moving on
The interesting thing about this release is that the pricing page is the least informative part of it. The list price did not move, so the whole delta is in behavior: thinking on by default, a 400 that did not exist before, a cache floor cut in half, and an effort ladder whose recommended entry point is the middle rung rather than the top one. Anthropic also published five benchmark rows where its new flagship loses, which is worth more than the nine where it wins.
If you are migrating, the work is small and specific. Change the model string, audit the routes that disable thinking, raise max_tokens where thinking is now implicit, delete the verification instructions you wrote for an older model, and sweep effort against your own evals rather than inheriting a setting. That is how to use Claude Opus 5 without paying for the top rung of a ladder that stops climbing.
References
- Anthropic. Introducing Claude Opus 5 — benchmark comparison table and effort-level cost curves. Retrieved July 2026 from anthropic.com/news/claude-opus-5
- Anthropic. What's new in Claude Opus 5 — behavior changes, new features, availability, and migration. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/models/whats-new-opus-5
- Anthropic. Pricing — token, cache, batch, and fast-mode rates. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/pricing
- Anthropic. Rate limits — per-model buckets and fast-mode limits. Retrieved July 2026 from platform.claude.com/docs/en/api/rate-limits
- Anthropic. Service tiers — Priority Tier model coverage. Retrieved July 2026 from platform.claude.com/docs/en/api/service-tiers
- Anthropic. Release notes — Claude Opus 5 launch entry dated July 24, 2026. Retrieved July 2026 from platform.claude.com/docs/en/release-notes/overview
Further reading
- reAPI. What is Claude Opus 4.8? reapi.ai/blog/what-is-claude-opus-4-8
- reAPI. Claude Opus 5 endpoint reference. reapi.ai/docs/claude-opus-5
- reAPI. Model catalog. reapi.ai/models
Autor
Kategorien
Weitere Beiträge
Veo 3.1 vs Seedance 2.0: Picking a Video Model in 2026
Picking Veo 3.1 vs Seedance 2.0 in 2026? Two very different bets in AI video. Capability, multi-shot, audio, resolution, and price with sourced numbers.
Best fal.ai Alternatives in 2026: 5 Options Compared
Looking for fal.ai alternatives in 2026? We compare Replicate, Together AI, RunPod, Hugging Face, and reAPI on model range, pricing, speed, and API design.
Best Venice.ai Alternatives in 2026: 5 Options Compared
Looking for Venice.ai alternatives in 2026? Compare OpenRouter, Together AI, DeepInfra, local Ollama, and reAPI on models, privacy, pricing, and API design.