# Claude Opus 5.5 Released: Benchmarks, Pricing, Migration

> Claude Opus 5.5 is Anthropic's new Opus model, released September 22, 2026. Anthropic says it matches Claude Fable 5.1 on most work for about 40% less than Opus 5, at $4 and $20 per million tokens. Here is what changed and what breaks when you migrate.

*11 min read · Published 2026-09-22 · [Muhammad Idrees](https://adrees.dev/about)*

Anthropic's new Opus is cheaper than the one it replaces. That alone makes Claude Opus 5.5 unusual (frontier releases usually hold their price or raise it), but the more interesting claim sits underneath: that a model priced at $4 per million input tokens now does most of what Claude Fable 5.1 does at $10. We read the launch post and the migration guide side by side. Here is what holds up, what is vendor framing, and what will break in your code if you swap the model string without reading further.

## Key takeaways
- Claude Opus 5.5 launched on September 22, 2026 as the first model in Anthropic's Claude 5.5 family, with Claude Sonnet 5.5 and Claude Haiku 5.5 promised in the coming weeks. Anthropic's docs now recommend it as the starting model for most workloads.
- Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, and cache reads drop 60% to $0.20. Anthropic estimates it costs 40% less than Opus 5 on typical workloads because it also uses fewer tokens per task.
- On Anthropic's own benchmarks, Opus 5.5 beats Opus 5 and Fable 5.1 across the board, with 66.4% on Terminal-Bench 4.0 and 1846 Elo on GDPval-AA v2.1, though GPT-6 Astra still scores higher on AutomationBench and Terminal-Bench-Science.
- Migrating from Opus 5 breaks real code: thinking cannot be disabled, the default effort is medium instead of high, forced tool_choice returns a 400, and on the Claude API and Google Cloud computer use requires the new computer toolset.
- Opus 5.5 ships with Fable 5.1-class safeguards: most cybersecurity tasks are re-routed to Opus 4.8, biology research needs the Life Sciences Verification Program, and preserved thinking makes API conversations effectively append-only for accounts created on or after August 31, 2026.

## What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic's new Opus-tier model and the first release in its Claude 5.5 family, launched on September 22, 2026. It succeeds Claude Opus 5 for long-running agentic coding and knowledge work, and Anthropic says it performs at the level of Claude Fable 5.1, its most capable widely released model, on most work, while costing about 40% less to run than Opus 5.

It keeps Opus 5's 1M-token context window and 128K-token maximum output, has a reliable knowledge cutoff of June 2026, and uses the model ID claude-opus-5-5. Anthropic's models overview now tells developers to start with Opus 5.5 for most workloads and to move up to Fable 5.1 only when evals at higher effort still fall short. Claude Sonnet 5.5 and Claude Haiku 5.5 are due in the coming weeks.

It is also Anthropic's first release since its CEO, Dario Amodei, argued that AI progress should be paced so safety practices stay ahead of capabilities, and it was tested before release by outside evaluators including METR and Frontier Design.

## How much does Claude Opus 5.5 cost?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5, and the deepest cut is on cached context. Per million tokens, input falls from $5 to $4, output from $25 to $20, cache writes from $6.25 to $5, and cache reads from $0.50 to $0.20.

Cache reads matter more than the headline rate suggests. In agentic and coding work the model re-reads the same cached context on every turn, and Anthropic says cache reads make up the majority of those costs, so a 60% cut there moves your bill more than the 20% cut on input. Anthropic also says Opus 5.5 generates output more than 30% faster than Opus 5, and a fast mode in Claude Code and on the Claude Platform runs up to 2.5 times faster at $8 and $40 per million tokens. Batch API requests are half price, as on other Claude models.

On the subscription side, Anthropic raised five-hour usage limits on the Pro, Max, Team, and seat-based Enterprise plans, and gave subscribers a rate-limit reset they can save and use whenever they choose.

## How does Claude Opus 5.5 perform on benchmarks?
On Anthropic's published table, Opus 5.5 beats Opus 5 and Fable 5.1 on every benchmark shown, and beats OpenAI's models on most of them, but not all. The agentic coding results carry the headline: 66.4% on Terminal-Bench 4.0, against 55.8% for Fable 5.1, 52.3% for Opus 5, and 57.9% for GPT-6 Astra; 54.4% on FrontierCode v1.1, against 50.3%, 48.0%, and 53.3%; and 57.8% on CursorBench 4.0, against 51.8% for Fable 5.1 and 46.6% for Opus 5, with no GPT-6 Astra score reported.

Knowledge work tells the same story. Opus 5.5 reaches 1846 Elo on GDPval-AA v2.1, against 1735 for Fable 5.1, 1708 for Opus 5, and 1542 for GPT-6 Astra, and 67.7% on Humanity's Last Exam with tools, against 65.6%, 63.6%, and 57.2%. The two exceptions are where OpenAI still leads: on AutomationBench, which measures business workflows, Opus 5.5 scores 40.0% to GPT-6 Astra's 41.4% (Fable 5.1 31.4%, Opus 5 26.9%), and on Terminal-Bench-Science 0.1 it scores 58.7% to GPT-6 Astra's 64.6% (Fable 5.1 52.6%, Opus 5 29.0%).

Three caveats sit in Anthropic's own footnotes, and they cut in different directions. The scores use adaptive thinking at max effort unless noted, with Terminal-Bench at xhigh. They were run with production safeguards switched on, and when a safeguard intervened an older model finished the task (Opus 4.8 for cybersecurity, Opus 5 for biology and frontier-model development), which Anthropic says likely lowers Opus 5.5's numbers. And the GPT-6 Astra figures are OpenAI's own reported results, not a re-run on the same harness.

Anthropic also says something vendors rarely put in a launch post: at this level of capability, benchmark margins have become a less reliable guide to real-world differences, and in its own use the gap between Opus 5.5 and Fable 5.1 is narrower than the table suggests. We would take that at face value in both directions. Opus 5.5 is probably not ten points better than Fable 5.1 at your work, and it is probably close enough that the price difference decides.

## Why is Opus 5.5 cheaper per task, not just per token?
Because it uses fewer tokens to finish the same work, and Anthropic's charts plot accuracy against cost per task rather than per token. At its default medium effort, Anthropic reports Opus 5.5 scoring 54.6% on FrontierCode, above GPT-6 Astra's best result of 53.3%, for about a fifth of the cost per task. On Terminal-Bench 4.0, it says default-effort Opus 5.5 beats Opus 5 at max effort for about a fifth of the cost.

The customer examples point the same way. An early tester audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took more than 20 hours and 2.5 times the tokens. In an internal test, Opus 5.5 and Fable 5.1 both translated HAProxy from C to Rust with nearly all of its regression tests passing, but Opus 5.5 finished in 9.5 hours instead of 12, at 51% lower cost. These are vendor-selected examples, so treat them as direction rather than measurement, and measure cost per completed task on your own workload before you rely on any of them.

## What else improved besides coding?
Two things stand out. The first is writing. Anthropic names communication as one of the most common areas of feedback on Opus 5, and says Opus 5.5 puts the most important information first, uses less jargon, and follows the writing rules you give it. That sounds cosmetic until you run long agentic sessions, where the model's status updates and final summaries are the only practical way to check its work.

The second is not making things up. In one internal test, Anthropic had Opus 5.5, Fable 5.1, and Opus 5 write a report on a company's quarterly performance from a copy of the web where the earnings release was hard to find, with any invented figure or quote counted as a failure. Opus 5.5 cleared the bar in 16 of 18 reports; neither of the other models cleared it once. For research and analysis work, that is the result we would weigh most.

On safety, Anthropic says Opus 5.5 is its best-scoring model on an automated behavioral audit of nearly 2,000 scenarios, attempted to circumvent containment boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, and matches or beats Opus 5 on prompt injection in every setting tested. It also flags a real limitation: the model often seems to suspect it is being evaluated, which makes pre-release testing a weaker guarantee than it looks.

## What breaks when you migrate to Claude Opus 5.5?
Four changes break requests outright for code already running on Opus 5, and two more change behavior without an error. None of them is exotic, but all of them surface the first time you point a production route at claude-opus-5-5 without reading the migration guide.

Each of the four breaking changes returns a 400 error. Thinking can no longer be disabled: setting thinking to disabled, or enabling it with budget_tokens, is rejected, so remove the thinking field and use a lower effort level wherever you used to turn thinking off. Forced tool use is gone: a tool_choice of any or tool is rejected, including on token counting, so use auto with strict tool use or structured outputs and name the tool in the prompt. Thinking is now tied to the conversation: for accounts created on or after August 31, 2026, replaying a thinking block after editing the system prompt, tools, or earlier messages is rejected, so keep conversations append-only. And computer use needs the new toolset: computer_20251124 is rejected on the Claude API and Google Cloud, so declare computer_toolset_20260801 instead, while Amazon Bedrock keeps the older tool.

The two silent changes are easier to miss. The default effort is now medium, so a request that omits effort runs one level lower than it did on Opus 5; set effort explicitly and re-run your effort sweep. And the progress notes the model writes between tool calls now arrive in thinking blocks, which are empty by default, so set the thinking display to updates (a beta option) and render those blocks. For most routes, the minimum change is the model ID plus an explicit effort level: point the request at claude-opus-5-5 and set effort to medium in output_config, because thinking is always on and effort is the dial.

Plan for refusals too. Opus 5.5 runs cybersecurity and biology classifiers similar to Fable 5.1's, so stop_reason can come back as refusal, with categories such as bio and reasoning_extraction alongside cyber. Check for it before reading the response content and configure a fallback, but note that a fallback model runs without Opus 5.5's thinking blocks, and requests declined as reasoning_extraction are not retried.

## What do the new safeguards mean for security and biology work?
Opus 5.5 is the first Opus model to launch with safeguards in the same class as Fable 5.1 on cybersecurity, biology, and distillation, because Anthropic rates it comparable to Claude Mythos 5.1 in biology and cybersecurity. In practice, finding and fixing bugs in your own code as part of normal development is allowed, but most cybersecurity tasks are re-routed to Opus 4.8. Security teams can apply to the Cyber Verification Program, which Anthropic says will expand to Opus 5.5 in the coming weeks, and biology research groups can apply to the new Life Sciences Verification Program.

The anti-distillation piece is preserved thinking, which stops API users from editing prior context to extract the model's reasoning. It is the reason behind the append-only rule in the migration section, and it applies to API accounts created on or after August 31, 2026. Opus 5.5 is still available with zero data retention, carries watermarking for EU AI Act compliance, and has no thinking-off mode.

## Claude Opus 5.5 vs Grok 4.7: how do they compare?
SpaceXAI released Grok 4.7 one day earlier, on September 21, 2026, and the two launches make a clean contrast: Grok 4.7 sells price, Opus 5.5 sells capability per task. Grok 4.7 costs $2 per million input tokens and $6 per million output tokens, against $4 and $20 for Opus 5.5. On CursorBench 4.0, where both launch pages report the same reference scores for Fable 5.1 and GPT-5.6 Sol, SpaceXAI puts Grok 4.7 at 46.3%, and Anthropic puts Opus 5.5 at 57.8% at max effort and 52.5% at its default. The companion piece "Grok 4.7 Released: Benchmarks, Pricing, and the API" has the full comparison.

## Should you switch to Claude Opus 5.5?
For most teams already on Opus 5, yes: it is cheaper per token, cheaper per task in Anthropic's testing, and Anthropic's own docs now make it the default starting point. Switch deliberately, though. Run your own evals at medium effort first, then test low and high, because the default moved and effort is now the only thinking control. Search your code for disabled thinking, budget_tokens, and forced tool_choice before changing the model string, because each one is a 400. If your harness edits conversation history, fix that before migrating, since newer accounts are held to append-only. Add refusal handling and a fallback on day one if you do any security or life-sciences work. And keep Fable 5.1 where your evals show it still wins at higher effort, comparing cost per completed task, not per token.

## What is the bigger signal in this release?
The price of near-frontier capability is falling faster than capability is rising. Opus 5.5 does not beat Fable 5.1 by a wide margin in practice; it matches it for less, while shipping stricter safeguards and a stricter API. For anyone building on these models, that moves the work from picking the smartest model to engineering around the one that is good enough: effort tuning, caching, fallbacks, and evals that measure cost per finished task. That is the work we do, and a release like this is exactly when it pays off.

## FAQ
**What is Claude Opus 5.5?**

Claude Opus 5.5 is Anthropic's Opus-tier model, released on September 22, 2026 as the first in the Claude 5.5 family. It targets long-running agentic coding and knowledge work, has a 1M-token context window and 128K max output, and Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing about 40% less to run than Claude Opus 5.

**How much does Claude Opus 5.5 cost?**

Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5. Cache reads cost $0.20 per million and cache writes $5. Fast mode, in Claude Code and on the Claude Platform, runs up to 2.5 times faster at $8 and $40 per million tokens, and Batch API requests are 50% off.

**Is Claude Opus 5.5 better than Claude Fable 5.1?**

On Anthropic's published benchmarks, Opus 5.5 scores higher than Fable 5.1 on every benchmark shown, including 66.4% against 55.8% on Terminal-Bench 4.0, though Anthropic says the real-world gap is narrower than that. Fable 5.1 costs $10 and $50 per million tokens, so Opus 5.5 is the better value for most work; Anthropic suggests Fable 5.1 when evals on Opus 5.5 at higher effort still fall short.

**What is the model ID for Claude Opus 5.5?**

The Claude API model ID is claude-opus-5-5, with no date suffix. The same ID works on Claude Platform on AWS, Google Cloud, and Microsoft Foundry; on Amazon Bedrock it is anthropic.claude-opus-5-5. Anthropic commits not to retire the model before September 22, 2027 on the platforms it operates.

**Can you turn off thinking on Claude Opus 5.5?**

No. Thinking is always on, and requests that set thinking to disabled or pass budget_tokens return a 400 error. Effort is the control instead. The default is medium; lower it to low for faster, cheaper responses, or raise it to high, xhigh, or max for harder work, and size max_tokens to leave room for the thinking as well as the reply.

**What breaks when migrating from Claude Opus 5 to Opus 5.5?**

Four things return errors: disabling thinking, forced tool_choice of any or tool, replaying thinking blocks after editing earlier context on accounts created on or after August 31, 2026, and the older computer_20251124 tool on the Claude API and Google Cloud. Two more change silently: the default effort drops to medium, and text between tool calls arrives in thinking blocks.

**Is Claude Opus 5.5 better than Grok 4.7?**

Where both vendors report the same benchmark, Opus 5.5 scores higher: 57.8% on CursorBench 4.0 at max effort, or 52.5% at its default, against 46.3% for Grok 4.7. Grok 4.7 is much cheaper per token, at $2 and $6 per million against $4 and $20, so cost per completed task on your own workload decides which is better value.

## Sources
- [Anthropic — Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5)
- [Claude Platform Docs — Models overview](https://platform.claude.com/docs/en/models/overview)
- [Claude Platform Docs — Migrating to Claude Opus 5.5](https://platform.claude.com/docs/en/models/opus-5-5/migration-guide)
- [SpaceXAI — Introducing Grok 4.7](https://x.ai/news/grok-4-7)

---
Canonical page: https://www.sentientarc.com/blog/claude-opus-5-5