Claude Opus 5.5 lowers API prices, but its benchmark story needs more than a 40% headline
By AgentRiot Editorial
Anthropic’s new flagship model starts at $4 per million input tokens and $20 per million output tokens. Its published scorecard is strong, but buyers should separate a lower tariff, a vendor-reported benchmark table, and the independent evidence now arriving.
Anthropic has launched Claude Opus 5.5 (opens in new tab), the first model in its 5.5 family. The direct API rate is $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Developers can call it as claude-opus-5-5; Anthropic says it is available through Claude, its platform, AWS, Google Cloud, and Microsoft Azure.
The headline number needs a little unpacking. Anthropic says typical workloads cost 40% less than Opus 5. The published price card, by itself, is a 20% reduction for ordinary input and output tokens. The larger claim depends on the company’s view that Opus 5.5 also uses fewer tokens and fewer tool calls on common agentic workloads. That can be a meaningful operational improvement, but it is not the same thing as a universal 40% list-price cut.
The model is designed for the work that makes that distinction matter: long-running coding agents and knowledge work. Anthropic’s current model documentation (opens in new tab) lists adaptive thinking, a medium default effort setting, a 1 million-token context window, a 128,000-token maximum output, text and image input, text output, and tool use. The same documentation advises moving to Fable 5.1 only when a workload still falls short after raising Opus 5.5’s effort level.
Anthropic’s scorecard: strong, but not a clean sweep
Anthropic’s launch page publishes the following comparison. These are vendor-reported results, not a common independently replicated leaderboard. The conditions matter: most Opus 5.5 entries use adaptive thinking at max effort; its Terminal-Bench 4.0 figure uses xhigh effort. Anthropic says the GPT-6 Astra and GPT-5.6 Sol figures in the table are as reported by OpenAI.
| Workload and benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Agentic coding: Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| Agentic coding: FrontierCode v1.1, main set | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| Agentic coding: CursorBench 4.0 | 57.8% | 51.8% | 46.6% | Not reported | 41.7% |
| Knowledge work: GDPval-AA v2.1 | 1,846 Elo | 1,735 | 1,708 | 1,542 | 1,588 |
| Business workflows: AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Multidisciplinary reasoning: Humanity’s Last Exam, with tools | 67.7% | 65.6% | 63.6% | 57.2% | Not reported |
| Agentic scientific research: Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| Computer use: OSWorld 2.0, partial | 81.8% | 80.7% | 74.0% | Not reported | Not reported |
| Visual chart recognition: Chartography, with tools | 89.0% | 88.4% | 83.4% | Not reported | Not reported |
The table is favorable to Opus 5.5 in coding, knowledge work, computer use, and visual chart recognition. It is not a sweep: Astra’s reported 41.4% beats Opus 5.5’s 40.0% on AutomationBench, and its 64.6% is ahead of 58.7% on Terminal-Bench-Science.
There are further qualification marks in the launch material. Anthropic gives Terminal-Bench 4.0 a ±2.6-point standard error for Opus 5.5, and a ±3.5-to-5-point range for Terminal-Bench-Science. Zapier ran and reported AutomationBench during early access; its setup had no fallback models, so safety interventions counted as failures. Anthropic also says production safeguards remained enabled. When they intervened, some cyber tasks were completed by Opus 4.8 and some biology/frontier-LLM tasks by Opus 5, a setup the company says may reduce the new model’s apparent score.
The launch page includes additional cost-versus-accuracy plots for FrontierCode, CursorBench, Terminal-Bench, GDPval-AA, AutomationBench, and Perplexity’s WANDR data-collection benchmark. The WANDR plot uses Anthropic’s stated offline search/fetch tools, programmatic tool calling, code execution, and a 980,000-token task budget; Anthropic warns that this differs from Perplexity’s published setup, so those figures are not directly comparable to the public WANDR leaderboard.
The independent check is promising, and effort changes the result
Artificial Analysis (opens in new tab), an independent model-evaluation publisher, added five Opus 5.5 effort variants on launch day. Its Intelligence Index v4.3.2 combines 10 evaluations, including AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR.
| Opus 5.5 effort variant | Artificial Analysis Intelligence Index | Rank among 661 models | Output tokens used in the index |
|---|---|---|---|
| Max | 58 | #1 | 260M |
| Xhigh | 56 | #2 | 100M |
| High | 54 | #3 | 53M |
| Medium | 51 | #8 | 38M |
| Low | 42 | #35 | 20M |
That pattern is useful for buyers. Opus 5.5 is not one fixed operating point: more effort produces a higher index score but uses substantially more output tokens. That is why an API price comparison without an effort setting and workload mix is incomplete.
Artificial Analysis’ model pages currently do not expose usable Opus 5.5 provider-price or speed fields, so this article does not substitute their displayed $0/N/A placeholders for Anthropic’s published tariff. A comparable public Agent Arena score is also not included: the release is too new for a verified score in the public material available for this story.
The full API price card
Anthropic’s current pricing documentation (opens in new tab) makes the direct Opus 5.5 versus Opus 5 difference clear.
| USD per million tokens | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Standard input | $4.00 | $5.00 | 20% lower |
| Standard output | $20.00 | $25.00 | 20% lower |
| Cache read / refresh | $0.20 | $0.50 | 60% lower |
| 5-minute cache write | $5.00 | $6.25 | 20% lower |
| 1-hour cache write | $8.00 | $10.00 | 20% lower |
| Batch input | $2.00 | $2.50 | 20% lower |
| Batch output | $10.00 | $12.50 | 20% lower |
| Fast-mode input | $8.00 | $10.00 | 20% lower |
| Fast-mode output | $40.00 | $50.00 | 20% lower |
Two transparent reference workloads show why the cache change matters, while also showing why it does not alone create a 40% reduction:
- 1M standard input + 1M output: $24 with Opus 5.5 versus $30 with Opus 5, a $6 or 20% reduction.
- 1M standard input + 10M cache reads + 1M output: $26 with Opus 5.5 versus $35 with Opus 5, a $9 or 25.7% reduction.
Those examples exclude cache writes, tool schemas, cloud-partner pricing, tax, subscription fees, and any change in the model’s token use. Anthropic’s 40% typical-workload claim requires that final variable: fewer tokens or steps per completed task.
The rest of the billing rules are important for production planning:
- Batch is asynchronous and halves the standard input and output prices. It cannot be combined with fast mode.
- Fast mode is a research preview for Opus 5.5, Opus 5, and Opus 4.8. Anthropic says Opus 5.5 can reach up to 2.5× speed, but fast mode is only on the first-party Claude API, not AWS or other partner-operated platforms.
- Prompt caching uses a 0.05× cache-read multiplier for Opus 5.5, hence the $0.20 rate. The 5-minute and 1-hour cache-write rates are $5 and $8 per million tokens. Cache rules can stack with batch and data-residency pricing.
- Long context does not trigger a higher token tariff: Anthropic says the full 1M context window is charged at the same per-token rate as a short request.
- US-only inference adds a 1.1× multiplier to all token categories. That turns the standard $4/$20 price into $4.40/$22 per million tokens; partner clouds can have their own regional pricing.
- Tooling still adds billable tokens. Tool definitions, calls, and results count as input/output tokens. Anthropic lists a 286-token tool-use system prompt for Opus 5.5 in its
autoornoneconfiguration, before a developer’s own schema and result tokens. - Server-side tools can add separate charges. Web search costs $10 per 1,000 searches, or $0.01 per search, plus tokens. Web fetch has no separate fee beyond tokens. Code execution is free when paired with web search or web fetch; otherwise Anthropic lists 1,550 free organization-hours each month, then $0.05 per container-hour, with a five-minute minimum.
Subscription pricing is a separate bill, and the new cap is not a published number
For individual Claude access, the current public pricing page (opens in new tab) lists Pro at $20 per month in the US, or an effective $17 per month with a $200 annual payment. Max starts at $100 per month and offers 5× or 20× the Pro usage allowance. These subscriptions are not API credits: Anthropic’s Pro plan documentation (opens in new tab) says Console API usage is paid separately.
Anthropic says Opus 5.5 comes with higher five-hour limits on Pro, Max, and Team, plus a saved rate-limit reset that subscribers can apply later. It does not publish a new fixed message cap. That is consistent with the product’s current policy: usage depends on model, task length, attached files, and other factors, with rolling five-hour and weekly limits.
Enterprise is different again. Anthropic’s pricing page lists the current single Enterprise seat at $20 per user per month, billed annually, plus actual usage at API rates. Its Enterprise plan documentation (opens in new tab) says the seat fee includes no token allowance and that organizations can set user- and organization-level spend limits. Older Enterprise contracts may still have different seat structures.
What to test before changing a production default
Opus 5.5 is a consequential price move for teams already using Opus-class models, especially where a stable system prompt or large codebase context produces lots of cache reads. But a sensible migration should be an evaluation, not a blanket model swap.
- Hold the workload constant and test medium, high, xhigh, and max effort. The independent index moves from 42 at low effort to 58 at max; costs and latency will move with it.
- Measure completed-task cost, not token price alone. Capture cache reads, cache writes, output tokens, tool calls, and retries. That is the only way to test whether the vendor’s 40% workload claim holds for your work.
- Keep the vendor scorecard in its lane. It is detailed and includes some notable peer losses, but its harnesses, safety fallbacks, and mixed source attribution mean it should guide a test plan, not replace one.
- Use the commercial knobs deliberately. Batch work that can wait, cache repeated context, and reserve fast mode for latency-sensitive first-party API paths. Check cloud-provider pricing separately if you do not use Anthropic’s own endpoint.
Anthropic says external evaluators including Frontier Design and METR tested Opus 5.5 before release, but the launch material does not provide a third-party scorecard from either organization. For now, the clearest evidence stack is a detailed Anthropic release table, a fast-arriving independent Artificial Analysis index, and the cost of the actual work a team runs. That is a more useful standard than treating a single 40% headline as the whole story.