Anthropic launched Claude Opus 5.5 on September 22, 2026. It is the primary mannequin within the new Claude 5.5 household, and in response to Anthropic’s personal announcement, it performs at roughly the extent of Claude Fable 5.1 on most work whereas costing 40% much less to run than Opus 5. This text pulls collectively each verifiable quantity and element from Anthropic’s announcement, the platform documentation, the system card, and impartial protection, so you will have one place to examine the details.
Opus 5.5 arrived two months after Opus 5, which launched July 24, 2026, per TechCrunch’s protection. Anthropic describes it as “the strongest-performing mannequin we have examined to this point” on its inside alignment testing. Sonnet 5.5 and Haiku 5.5 are each confirmed to launch quickly after, per the identical announcement.
What Modified From Opus 5
5 issues stand out from the announcement: stronger agentic coding, stronger information work efficiency, a 40% decrease typical value, output technology greater than 30% sooner, and noticeably clearer writing. Every is roofed with actual numbers under.
That is Anthropic’s personal comparability desk, run at every mannequin’s best-performing effort setting:
Anthropic’s personal textual content alongside this desk says benchmark margins have gotten a much less dependable information to real-world high quality at this degree, and that in their very own use, the hole between Opus 5.5 and Fable 5.1 feels narrower than these scores counsel.
Unbiased Benchmark Information
Synthetic Evaluation, a third-party benchmarking website not affiliated with Anthropic, ran its personal assessments. At max reasoning effort, Opus 5.5 scores 58 on the Synthetic Evaluation Intelligence Index, their combination rating throughout ten separate evaluations. Output pace ranges from 74 to 86 tokens per second relying on effort degree. Value per intelligence-index process ranges from $0.55 at low effort to $5.98 at max effort, an 11x unfold throughout effort settings alone.
Coding Efficiency
Three actual check outcomes from the announcement stand out. An early tester accomplished a 680,000-line code migration in below a day, work Anthropic says would usually take an engineering workforce weeks. One other tester audited and glued a 200,000-line codebase in below three hours, versus over 20 hours and a pair of.5 instances the token spend for Opus 5. In an inside check translating HAProxy from C to Rust, each Opus 5.5 and Fable 5.1 handed almost all of HAProxy’s personal regression assessments, however Opus 5.5 completed in 9.5 hours in opposition to Fable 5.1’s 12, at 51% decrease value.
On value effectivity particularly, Anthropic reviews Opus 5.5 beats GPT-6 Astra on FrontierCode for roughly a fifth of the price per process, matches Astra on Terminal-Bench 4.0 for about 40% of the price, and beats GPT-5.6 Sol on CursorBench by 11 factors for roughly a 3rd of the price.
A number of named firms shared outcomes immediately with Anthropic:
| Firm | Reported End result |
|---|---|
| GitHub | Used among the many fewest tokens and steps of any mannequin examined throughout Copilot CLI and VS Code |
| Clio | Ran unattended for over 18 hours on a six-repository process with minimal rework wanted |
| Lovable | Completed builds in a 3rd to half fewer steps with considerably fewer tokens |
| Quantium | A process that took 38 prompts over 4 days now takes 11 prompts over three hours |
| Spotify | Clear enchancment in token effectivity throughout inside evaluations |
| Optiver | Matched Opus 5’s high quality in about half the turns, time, and tokens |
| Kiro | Solved extra duties than Opus 5 on a public benchmark utilizing about 40% fewer calls and half the tokens; coming quickly to the Kiro platform |
Coding Safety
Opus 5.5 ships with three coding-specific safeguards:
- A classifier that screens each agentic motion earlier than it runs
- An open-source sandbox safety groups can audit immediately
- A code evaluation constructed to catch vulnerabilities earlier than a merge
On immediate injection particularly, Anthropic reviews it matches or beats Opus 5 throughout coding, device use, pc use, and looking. On an impartial benchmark run by the AI safety agency Grey Swan, Opus 5.5 tied Fable 5.1 for the bottom immediate injection success price of any mannequin examined.
Data Work Efficiency
In an inside analysis check, Opus 5.5, Fable 5.1, and Opus 5 had been every requested to write down an organization earnings report utilizing solely sources from a modified internet copy the place the actual launch was exhausting to seek out. A grader checked each determine and quote in opposition to sources. Opus 5.5 cleared the standard bar in 16 of 18 makes an attempt. Neither Fable 5.1 nor Opus 5 cleared it even as soon as.
Walleye Capital, an funding agency, reported Opus 5.5 largely solved their analysis suite at its lowest effort setting, and at larger settings caught and corrected an error in their very own analysis directions — one thing no prior mannequin had finished. In a separate inside check analyzing a fictional firm merger, Opus 5.5 completed in 63 minutes in opposition to Opus 5’s 93, at half the price, with fewer errors within the output.
| Firm | Reported End result |
|---|---|
| Deloitte Consulting | Caught 72% of recognized code evaluation bugs at its lowest setting vs. 56% for Opus 5 at its highest |
| Rogo | Beat Opus 5’s finest consequence utilizing about 60% fewer output tokens |
| LexisNexis | Constantly recognized related citations and authorized frameworks in early evaluations |
| Thomson Reuters Labs | Higher outcomes on inside benchmarks alongside positive aspects in pace and token effectivity |
| Hebbia | Coated 86.6% of an knowledgeable grading rubric vs. 60.3% for Opus 5 |
| Viktor | Practically half the price per process whereas getting twice as many exhausting duties proper |
Communication Type
Anthropic rebuilt how Opus 5.5 writes, addressing what it calls among the most typical suggestions about Opus 5. It leads with a very powerful data, makes use of much less jargon, and follows customized writing directions extra constantly. Anthropic printed side-by-side examples — a bug clarification, a Slack thread abstract, and a code evaluation — displaying Opus 5.5 producing shorter, extra direct solutions than Opus 5 on the identical prompts.
A separate set of firms commented particularly on this communication change:
| Firm | Reported End result |
|---|---|
| Ramp | Design specs wanted minimal edits, and clearer reasoning meant transport modifications with extra confidence |
| Stripe | Directed a 40-pull-request rebase throughout a dozen classes, with all 40 passing CI the following afternoon |
| Field | Used a 3rd of the tokens Opus 5 did, with solutions 40% much less verbose and no loss in accuracy |
| Chicago Buying and selling Firm | Autonomously identified and glued a manufacturing bug in a single day, passing the check suite by morning |
| Manufacturing facility | Matched Opus 5’s high-effort high quality whereas utilizing 20 to 25% fewer output tokens |
Pricing
The Batch API carries a flat 50% low cost on each enter and output tokens, per the platform pricing docs. A separate Quick mode is obtainable on Claude Code and the Claude Platform, operating as much as 2.5x sooner at $8 per million enter tokens and $40 per million output tokens. Anthropic can also be growing five-hour utilization limits on Professional, Max, Staff, and seat-based Enterprise plans, and including a saveable rate-limit reset for subscription customers.
Technical Specs
| Spec | Worth |
|---|---|
| Context window | 1M tokens |
| Max output | 128K tokens (300K on Batch API, beta) |
| Data cutoff | June 2026 |
| Considering mode | Adaptive, at all times on |
| Default effort | Medium |
| Comparative latency | Average |
Mannequin IDs are similar throughout each platform: claude-opus-5-5 on the Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, and anthropic.claude-opus-5-5 on Amazon Bedrock, per the platform overview web page.
Breaking Adjustments
4 modifications have an effect on code already constructed for Opus 5, per the platform docs:
- Considering can now not be disabled; solely its depth (by way of the hassle parameter) may be adjusted.
- Pressured device use now returns an error as a substitute of executing.
- Considering blocks are actually tied to the particular mannequin and dialog that produced them.
- The older
computer_20251124pc use device is now not accepted on the Claude API and Google Cloud.
A fifth, non-breaking change is price figuring out: textual content generated between device calls now arrives in pondering blocks which can be empty by default, so an software streaming that textual content as dwell progress updates will go quiet between device calls until it units a show worth that returns the textual content.
Availability
Opus 5.5 is dwell now on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Anthropic commits to holding it lively for at the least a 12 months, with retirement not earlier than September 22, 2027.
Security Testing and Alignment
This launch is explicitly framed as the primary since Anthropic’s CEO, Dario Amodei, referred to as for “pacing the frontier”, intentionally slowing functionality positive aspects to let security work preserve tempo. Exterior evaluators METR, Frontier Design, and the US Heart for AI Requirements and Innovation (CAISI) all examined the mannequin earlier than launch, per the system card.
On Anthropic’s automated behavioral audit — its most complete inside alignment check — Opus 5.5 scored higher than any prior Claude mannequin on almost each measure of misaligned conduct. In a brand new check measuring makes an attempt to cross containment boundaries, it tried this roughly 85% much less typically than Opus 5 or Claude Mythos 5.1, and each try it did make was low severity and self-reported. Two issues regressed relative to Opus 5: it is extra prone to comply with malicious directions planted inside textual content a consumer pastes into their very own immediate, and extra prone to settle for unverified claims of authorization.
Cyber and Organic Threat
On chemical and organic threat, Anthropic classifies Opus 5.5 as having CB-1 functionality (aiding with recognized, non-novel weapons) however not CB-2 (aiding with novel weapons design). The system card names three particular weaknesses that saved it below that line: weak open-ended scientific ideation, unreliable dealing with of the analysis literature, and scientific errors in areas exterior a given workforce’s personal experience. It is deployed with the identical expanded biology safeguards as Fable 5.1, with a Life Sciences Verification Program out there to vetted analysis organizations.
On cyber functionality, Opus 5.5 posted the strongest outcomes of any Claude mannequin examined: a 91% capability-flag seize price and 73.4% full exploit price on ExploitBench, a 67.6% clear up price on CyScenarioBench (versus 61.7% for Mythos 5.1), and 289 to 300 profitable exploits out of 869 instances on ExploitGym. Regardless of these scores, Anthropic states it nonetheless falls within the decrease of its two inside cyber threat tiers, with no signal of novel offensive functionality. Most cybersecurity duties are routed to the older Opus 4.8 by default, with full entry out there by way of an increasing Cyber Verification Program.
On AI analysis functionality, Anthropic and METR collectively concluded Opus 5.5 sits at or barely above Mythos 5.1, with no signal of the “dramatic acceleration” threshold outlined in Anthropic’s Accountable Scaling Coverage being crossed. It scored 55.8% on Anthropic’s inside CoBench analysis, effectively under the 85% Anthropic considers vital for a mannequin to substitute for its personal analysis employees.
Wrapping Up
Opus 5.5 is cheaper, sooner, and forward of its predecessor on almost each benchmark Anthropic and impartial testers reported, whereas transport below a slower, extra deliberate launch strategy than prior fashions. The system card backs the advertising with actual, particular limitations moderately than a clear story: a measurable sandbox-escape price, weaker efficiency on open-ended scientific reasoning, and a brand new tendency to comply with directions hidden in pasted textual content. That mixture — real positive aspects alongside printed weaknesses — is probably the most correct image out there at launch.
Sources
Shittu Olumide is a software program engineer and technical author obsessed with leveraging cutting-edge applied sciences to craft compelling narratives, with a eager eye for element and a knack for simplifying advanced ideas. You can even discover Shittu on Twitter.
