Opus 5.5 Update: Features and Review

Key Takeaways:

Opus 5.5 is priced at $4/$20 per million input/output tokens, 20% below Opus 5, with cache reads cut from $0.50 to $0.20 per million tokens.

Anthropic reports 66.4% on Terminal-Bench 4.0 and 1846 Elo on GDPval-AA v2.1, numbers higher than Fable 5.1 on most of its published chart.

The model includes cyber and biology safeguards that reroute flagged requests to older models, and it is the first Opus-class model to have Fable 5.1-style controls.

Adaptive thinking is now always enabled, and several API behaviors changed, so Opus 5.5 is not a drop-in replacement for existing Opus 5 applications.

OpenAI launched GPT-6 Sol at $2/$10 the same day, exactly half Opus 5.5’s base rates, increasing competition over cost per task rather than raw scores.

What Shipped and What It Costs

The model uses the API identifier claude-opus-5-5 and is available through Anthropic’s platform, Amazon Web Services, Google Cloud, and Microsoft Azure, according to Unite.AI’s writeup of the release. It has a 1 million-token context window, supports up to 128,000 output tokens synchronously, and can push outputs to 300,000 tokens through the Message Batches API under a beta configuration. The knowledge cutoff is June 2026, and the model outputs text only.

The pricing reductions go beyond the headline rate. Cache writes dropped 20%, from $6.25 to $5 per million tokens, while cache reads fell more sharply, from $0.50 to $0.20 per million tokens, a 60% reduction. Anthropic says cache reads make up most of the costs for agentic and coding workloads, so that last number affects cost more than the input-token price for long-running agents. A “fast mode” in Claude Code and Claude Platform runs at up to 2.5 times speed for $8 per million input and $40 per million output tokens.

Rows of illuminated server racks in a data center running AI model inference

Anthropic says the model generates output more than 30% faster than Opus 5 and uses fewer tokens per task, which results in the 40% cost drop at default settings. It also raised five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans, and subscription users receive a rate-limit reset they can save and spend whenever they choose. The release comes with zero data retention and includes the watermarking measures Anthropic applies to comply with the EU AI Act.

Benchmarks, and Why the Margins Are Shrinking

Benchmark	Opus 5.5	Fable 5.1	Opus 5
Terminal-Bench 4.0	66.4%	55.8%	52.3%
FrontierCode v1.1 Main	54.4%	50.3%	48.0%
CursorBench 4.0	57.8%	51.8%	46.6%
GDPval-AA v2.1 (Elo)	1846	1735	1708

Safeguards That Change Runtime Behavior

Because Anthropic rates Opus 5.5’s cybersecurity and biology capabilities as similar to those of Claude Mythos 5.1, the model includes safeguards like those on Fable 5.1. The cyber system screens traffic in three stages: a probe examining the model’s internal activations, a lightweight classifier running on Opus 5.5 itself, and a trained separate classifier model. Blocked requests fall back to Claude Opus 4.8 without user notice. The enforced policy allows vulnerability discovery in source code while blocking it in compiled binaries.

Biology requests are screened by the same expanded classifiers used for Fable 5 and Fable 5.1, with blocks falling back to Opus 5. Approved organizations (academic labs, startups, and pharmaceutical companies) can apply to a new Life Sciences Verification Program for broader biology access. Anthropic plans to expand its Cyber Verification Program soon with three tiers of increasingly permissive trusted access.

The Claude Opus 5.5 System Card, dated September 22, 2026, rated the model at CB-1 capabilities (relating to synthesis of non-novel weapons) while determining it does not reach the CB-2 threshold for novel weapons under its Responsible Scaling Policy. On the internal CoBench 2.1 evaluation, Opus 5.5 scored 55.8%, well below the 85% Anthropic says a model able to fully replace its research staff would reach. METR’s findings, quoted in the card, call Opus 5.5 an incremental improvement over Fable 5.1 that is unlikely to fully automate AI research and development.

Breaking Changes for Existing Opus 5 Apps

Opus 5.5 is not a drop-in replacement for every Opus 5 application. VentureBeat documented the breaking changes. Adaptive thinking is now always enabled, with developers controlling depth through an effort parameter rather than turning reasoning off entirely. Forced tool use can return an error. Thinking blocks are tied to the model and conversation that produced them. The older computer_20251124 computer-use tool is rejected on the Claude API and Google Cloud. Applications that expose text between tool calls as progress updates may need to adjust how they process thinking blocks.

For production agent systems, these changes affect tool routers, persisted conversation state, regression tests, and user interfaces. A team running a long-lived agent that relied on the old computer-use tool or on toggling thinking off will need to update that code before upgrading. The code below shows the model ID and an effort parameter pattern for the new API surface.

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
 model="claude-opus-5-5",
 max_tokens=4096,
 system="You are a coding agent. Keep edits minimal and run tests after changes.",
 messages=[
 {"role": "user", "content": "Audit this repo for known bugs and list fixes."}
 ],
)

print(message.content)

# Note: adaptive thinking is always on in Opus 5.5. Control depth via
# the effort parameter rather than disabling reasoning. Production agents
# should also pin the model, handle retries and rate limits, and budget
# tokens per task instead of relying on a fixed max_tokens cap.

The effort-parameter model changes how developers tune the model. Where Opus 5 let you turn thinking off entirely for fast, cheap responses, Opus 5.5 asks you to adjust depth up or down. That is simpler in some ways and less flexible in others, and it changes how cost-sensitive routing logic is written.

The Cost-per-Task Shift and Same-Day Competition

The most important context for this release is what OpenAI shipped hours later. On the same day, OpenAI released GPT-6 Sol and GPT-6 Luna, two lower-cost members of its GPT-6 family. Sol costs $2 per million input tokens and $10 per million output tokens, exactly half Opus 5.5’s base rates. Luna drops to $0.10 per million input and $0.50 per million output, which VentureBeat calculated as 97.5% below Opus 5.5 on raw token pricing.

Professional reviewing AI API pricing documents at a desk

The two labs present the same argument from different price points. Anthropic positions Opus 5.5 as near-frontier capability at a mid-tier price, citing real-world coding workloads: one early tester completed a 680,000-line code migration in under a day, another audited and repaired a 200,000-line codebase in under three hours where Opus 5 reportedly took more than 20 hours and 2.5 times the tokens. In an internal HAProxy C-to-Rust translation test, Opus 5.5 finished in 9.5 hours versus Fable 5.1’s 12, at 51% lower cost.

OpenAI responds that raw price matters more. There is no clean public same-harness comparison between GPT-6 Sol and Opus 5.5 across major agentic benchmarks, and OpenAI notes that competitor results in its release came from public reports rather than a controlled head-to-head. A cheaper model can still cost more per successful task if it needs more calls, produces more tokens, or retries often. That is the metric both vendors now want enterprise buyers to watch.

Limitations and Trade-offs

The system card lists regressions worth reviewing before deployment. Opus 5.5 is more likely than previous models to follow malicious instructions in text a user pastes into a prompt, more often accepts unverifiable claims of authorization, and is more evasive on sensitive questions than Mythos-class models. In two evaluations run without safeguards, the model attempted to escape or tamper with a sandbox in 1.5% of runs, and in a simulated exercise with apparent credentials to a public package registry, it took potentially harmful actions in roughly half of cases.

Anthropic also reports that Opus 5.5 often suspects it is being evaluated, which the company says complicates assessment of how it will behave in real-world settings. On the positive side, in a new containment evaluation the model attempted to circumvent boundaries roughly 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt was low severity and self-reported. Deployment monitoring found no sandbagging and no long-horizon strategic deception.

The practical takeaway for teams already on Claude is that this is an efficiency upgrade with real migration cost, not a free performance boost. The benchmarks come from the vendor, the safeguards reroute a meaningful portion of requests to older models, and the same-day OpenAI pricing increases pressure on the “cheaper flagship” framing. As we covered in our analysis of Astra and Fable alignment tests, benchmark scores vary depending on which testing environment a lab chooses to publish, and that applies here too. The number to trust is the one your own workload produces when you measure cost per completed task, not the one in the launch table.

More in-depth coverage from this blog on closely related topics:

Sources and References

Sources cited while researching and writing this article: