Anthropic released Claude Opus 5.5 on September 22, 2026, two months after Opus 5, and cut the sticker price at the same time it raised what the model can actually do. Input tokens now cost $4 per million and output tokens $20 per million, down from Opus 5's $5 and $25. Cache reads dropped further, to $0.20 per million tokens, from $1.00 on 5-minute cache writes priced at $5 per million and 1-hour writes at $8. Anthropic frames the combined effect, a lower sticker price plus a model that needs fewer tokens to finish the same task, as roughly 40% lower cost to run on typical workloads. That is a blended figure, not a flat discount: input and output token prices dropped 20%, cache reads (which dominate the bill on any long agentic run) dropped 60%, and output now generates about 30% faster, and Anthropic nets those together rather than presenting one clean percentage off the sticker price.
The performance side moved just as much. Opus 5.5 scored 66.4% on Terminal-Bench 4.0, a benchmark that scores an autonomous agent on complex, multi-step engineering work carried out inside a real command-line environment, up from Opus 5's 52.3%. On Anthropic's own comparison set, agentic coding (FrontierCode v1.1: 54.4% versus Opus 5's 48.0%), IDE-style coding (CursorBench 4.0: 57.8% versus 46.6%), and business and knowledge work (GDPval-AA v2.1: an Elo of 1846 versus 1708), Opus 5.5 also edges out the larger, pricier Fable 5.1 model on every one of those four numbers, despite costing 60% less per output token than Fable 5.1's $50 per million. Anthropic's own framing of the model is more specific than a blanket capability claim: on what it calls its most comprehensive automated behavioral audit, the company describes Opus 5.5 as "the strongest-performing model we've tested to date."
The model is live now, on the Claude Platform at claude.ai, in Claude Code, through the Claude API under the model ID claude-opus-5-5, and through Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry on Azure. It carries a 1 million token context window and a 128,000 token maximum output, the same ceiling as Opus 5 and Fable 5.1, according to Anthropic's own model documentation. Sonnet 5.5 and Haiku 5.5 are not out yet; Anthropic says both are coming in the following weeks, which means anything built today is working with Anthropic's cheapest current access to this generation's coding and agentic strength. What follows are ten specific things worth starting on this week, grounded in what Anthropic actually published about where the model is strong: agentic coding, code migration and auditing, computer use, financial analysis, business workflows, frontend development, and long-horizon research.
10 Things Worth Building With Opus 5.5 Right Now
1. An Autonomous Codebase Migration Agent
Large, mechanical code migrations, framework upgrades, language ports, dependency overhauls, are exactly the kind of long, multi-step agentic work Opus 5.5 was built for. Anthropic's own example is a 680,000-line code migration completed in under a day, a scale of task that would previously have needed a team working for weeks under manual review. The model's 66.4% Terminal-Bench 4.0 score, a benchmark built around multi-step engineering work in a real terminal, is the direct evidence that it can hold a plan together across the hundreds of tool calls a migration like that requires. What makes this newly buildable by one person, rather than at Opus 5's pricing, is the token math: a migration agent spends most of its budget on cache reads and repeated context as it re-scans a repository, and those got 60% cheaper, so a job that would have cost real money to run start to finish on Opus 5 now clears the bar of being worth billing a client for. The tradeoff worth planning for before shipping one of these: a model doing this much of a migration alone raises the stakes on what your review step actually catches.
2. A Legacy Language Porting Service
Adjacent to migrations but distinct enough to be its own product: a small, productized service that takes a client's legacy C, COBOL, or old-Java codebase and ports it to a modern language, with a human reviewing the diff rather than writing it. Anthropic cites its own real-world test of this exact task: translating the HAProxy codebase from C to Rust in 9.5 hours, at 51% lower cost than the comparable run on Opus 5. A working port finished in under half a working day, at roughly half the compute cost of the previous model, is what turns this from an interesting demo into a service a two- or three-person shop can actually price and sell. Because the pitch depends entirely on beating what a human contractor would charge for the same port, a margin that was thin at Opus 5's pricing becomes comfortably workable at Opus 5.5's.
3. A Web App Performance Auditor
Point an agent at a live web app with computer-use access, a real browser it can click through, and have it profile load time, isolate the actual bottleneck, patch it, and re-measure, on a schedule, without a human starting each run. Anthropic reports Opus 5.5 succeeding on 39 of 40 web app load-time optimization tasks in testing, a materially higher hit rate than Opus 5 managed against the same set, where Anthropic notes the improvements were smaller. Computer use and frontend and web development are both strengths Anthropic names directly for this release. Running an agent like this continuously, checking a client's site after every deploy rather than once a quarter, only pencils out at a price where each run costs cents rather than dollars, which is what the 20% cut in per-token pricing plus faster output generation gets a solo developer to.
4. A Continuous Codebase Security and Tech-Debt Auditor
A different long-horizon use of the same coding strength: an agent that reads an entire codebase end to end, flags security issues, dead code, and structural debt, and produces a prioritized report, on a recurring basis rather than a one-time engagement. Anthropic's cited benchmark here is concrete: a 200,000-line codebase audit completed in under three hours, versus more than 20 hours for the same audit on Opus 5. That is not a marginal speedup, it is the difference between an audit that fits inside a single working session and one that needs to be scheduled as a multi-day engagement. For a freelancer or small shop, that turns a security and tech-debt audit from a scoped, expensive one-time project into something closer to a recurring, affordable subscription a client keeps paying for.
5. A Computer-Use Agent for One Specific Back-Office Workflow
Not a general-purpose office assistant, a narrow one: an agent that logs into one specific web tool a small business already uses, a supplier portal, a benefits admin site, a shipping dashboard, and does one repetitive task inside it correctly, every time, on its own schedule. Computer use is one of the strengths Anthropic names directly for this release. The reason this is worth building now rather than six months ago is cost of ownership: an agent that runs a browser session daily or hourly racks up token usage fast, and a 20% drop in the underlying token price plus cheaper cache reads is the difference between that workflow being profitable to automate for a client paying a modest monthly fee and it not being worth building at all.
6. A Financial Analysis Agent for Solo Operators and Small Funds
Anthropic names financial analysis and business workflows as specific strengths of this release, and the benchmark numbers back it: on GDPval-AA v2.1, built around real business and knowledge work, Opus 5.5 scored an Elo of 1846, ahead of Opus 5's 1708 and even ahead of Fable 5.1's 1735. That is a case for building an agent that pulls a company's filings or a fund's portfolio data, drafts a memo or a model, and flags what changed since the last run, work that used to require either a junior analyst's time or a much more expensive model call. Running that kind of analysis weekly for a handful of companies or a small portfolio through a slower, pricier model would eat into the margin an independent analyst actually needs to keep; at $4 input and $20 output per million tokens, the compute cost stops being the thing that decides whether the service is worth offering.
7. A Frontend Generator for One Specific Niche
Not another general app builder, a tool scoped to one vertical: generating fully working storefronts for a specific product category, or dashboards for a specific SaaS niche, from a structured brief rather than a blank prompt. Opus 5.5 scored 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0, both up sharply from Opus 5's 48.0% and 46.6%, and both ahead of Fable 5.1's 50.3% and 51.8% as well. Frontend and web development is one of the strengths Anthropic calls out by name for this model. A vertical frontend generator lives or dies on how many attempts it needs before the output is shippable without rework, and a jump of this size on the two benchmarks closest to that exact skill is what makes a narrow, opinionated generator, not a general-purpose one, realistic to build and charge for now.
8. A Long-Horizon Research Agent
Long-horizon research tasks are one of the strengths Anthropic lists for Opus 5.5 directly, and the model's 1 million token context window means an agent like this can hold an entire set of source documents, filings, or a competitor's public materials in view across a multi-hour run instead of re-fetching and re-summarizing them turn by turn. Pair that with output generation running about 30% faster than Opus 5, and a research pass that used to take an afternoon to finish a competitive teardown or a literature review now finishes in a fraction of that wall-clock time, at a lower cost per run. A one-person research operation that could previously only afford to run this kind of deep pass occasionally can now run it as a standing, always-on queue against a running list of questions.
9. An Internal Business-Process Automation Agent
Every small business has at least one process that is well-defined enough to automate but too idiosyncratic for off-the-shelf software: renewing a specific set of vendor contracts, reconciling a particular report against a particular spreadsheet, routing a particular kind of customer request. Business workflows are named directly among Opus 5.5's strengths, and the model's coding ability means an agent like this can be built as real, testable code rather than a fragile prompt chain that breaks the first time an exception shows up. The catch with any agent like this is that it needs to remember what happened last time, which vendor renewed on what terms, which exception was approved and why, across sessions that might be days or weeks apart, and that is a memory problem the model itself does not solve, no matter how cheap or capable it gets.
10. A Terminal-Native DevOps and Incident-Response Copilot
Terminal-Bench 4.0 exists specifically to measure how effectively an autonomous agent executes complex, multi-step engineering tasks inside a real command-line environment, and Opus 5.5's 66.4% on it, well ahead of Opus 5's 52.3% and Fable 5.1's 55.8%, is the most direct evidence available that this model can be trusted with a live terminal session. That makes it a strong fit for a copilot that watches an on-call channel, reproduces an incident in a real shell, checks logs and recent deploys, and proposes or executes a fix, a job that depends entirely on an agent not losing the thread across dozens of terminal commands in a row. For a small engineering team without a dedicated on-call function, the price drop is what makes leaving an agent like this running continuously, rather than invoking it only during a fire, something they can actually justify on a monthly bill.
Opus 5.5 vs. Opus 5 vs. Fable 5.1, at a Glance
The numbers that matter most for deciding what to build on are price, the Terminal-Bench 4.0 score Anthropic treats as its headline agentic-coding measure, and context window, since several of the ideas above depend on all three lining up at once.
| What You're Comparing | Claude Opus 5.5 | Claude Opus 5 | Claude Fable 5.1 |
|---|---|---|---|
| Price per million tokens (input / output) | $4 / $20 | $5 / $25 | $10 / $50 |
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 52.3% | 55.8% |
| Context window / max output | 1M tokens / 128K | 1M tokens / 128K | 1M tokens / 128K |
| Best fit for | Everyday agentic coding, migrations, computer use, and business workflows at scale | Existing deployments Anthropic now recommends migrating to 5.5 | The hardest reasoning and research tasks where accuracy matters more than cost |
Where MemX Fits: The Memory Problem Opus 5.5 Doesn't Solve
None of this changes what a model is: something that starts fresh, or close to it, every session. Idea 9's business-process agent above is a clean example of the gap. A vendor-renewal or exception-handling agent is only as useful as its memory of what happened the last ten times it ran, and neither Opus 5.5's larger context window nor its lower price does anything to persist that across separate sessions on its own; a fresh session still starts blind to what the last one decided. This is the specific, narrow problem a private, running memory layer like MemX is built to help with: a place to keep the decisions, exceptions, and context an agent or its operator needs to recall next time, kept private by architecture and usable across whichever combination of ChatGPT, Claude, and Gemini the underlying build actually relies on. It does not make an agent smarter or cheaper to run. It just means the fifth run of a workflow like idea 9 does not start from zero the way the first one did.
01What is Claude Opus 5.5 and when did Anthropic release it?
Claude Opus 5.5 is Anthropic's coding- and agent-focused model, released September 22, 2026, about two months after Opus 5. It is priced at $4 per million input tokens and $20 per million output tokens, and Anthropic describes it as built for long-running agentic coding and knowledge work.
02How much cheaper is Opus 5.5 than Opus 5, and what does the 40% figure actually mean?
Sticker price alone dropped 20% on input and output tokens, from $5/$25 to $4/$20 per million, and cache reads dropped 60%, to $0.20 per million. Anthropic's roughly 40% lower cost claim for typical workloads combines that price cut with the model needing fewer tokens to finish the same task and generating output about 30% faster, so it is not simply a flat 40% discount on the same bill.
03Is Claude Opus 5.5 better than Fable 5.1?
On the specific benchmarks Anthropic published alongside the release, agentic coding, IDE-style coding, and business and knowledge work, Opus 5.5 scored higher than Fable 5.1 while costing 60% less per output token. Fable 5.1 remains Anthropic's broader flagship model for the hardest reasoning tasks; Opus 5.5 is positioned specifically for agentic coding and long-running knowledge work at a lower price.
04What is Claude Opus 5.5's context window?
1 million tokens, with a maximum output of 128,000 tokens per request, the same limits as Opus 5 and Fable 5.1, according to Anthropic's own model documentation.
05Where can developers access Claude Opus 5.5?
On the Claude Platform at claude.ai, in Claude Code, through the Claude API under the model ID claude-opus-5-5, and through Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry on Azure. Sonnet 5.5 and Haiku 5.5 have not shipped yet; Anthropic says both are coming in the following weeks.
