Anthropic says Claude Opus 5.5 does Fable-level work for less than Opus 5 costs to run. Code already running on Opus 5 can break when you switch, though: Anthropic lists four breaking API changes, plus a fifth change that quietly alters what your app receives. Here's what changed, who should move, and a one-week upgrade plan.
Claude Opus 5.5 is Anthropic's newest Opus model, released September 22, 2026, as the first model in the Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. Before you swap model IDs, fix four breaking changes: thinking can't be disabled, forced tool use returns an error, thinking blocks are tied to the model and conversation, and the older computer_20251124 tool is rejected on the Claude API and Google Cloud. Opus 5 has no announced retirement date, so you can test first and move on your own schedule.
claude-opus-5-5), with a 1M-token context window, always-on adaptive thinking, and a default effort of medium.medium and higher effort, fix the breaking changes, and stage the rollout with Opus 5 pinned as a fallback.Every fact below comes from Anthropic's announcement or the Claude Opus 5.5 model page, checked September 23, 2026.
| Item | Claude Opus 5.5 |
|---|---|
| Release date | September 22, 2026 |
| Model ID | claude-opus-5-5 (anthropic.claude-opus-5-5 on Amazon Bedrock) |
| Family | First model in the Claude 5.5 family |
| Context and output | 1M-token context; 128K max output; 300K output on the Batch API with a beta header |
| Thinking and default effort | Adaptive thinking, always on; default effort medium (Opus 5 defaulted to high) |
| Price vs. Opus 5 | $4/$20 per million input/output tokens, down from $5/$25 |
| Speed | Output "more than 30% faster" than Opus 5, per Anthropic |
| Platforms | Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry |
| Retirement commitment | Not sooner than September 22, 2027 |
| Coming next | Claude Sonnet 5.5 now available ($2 input / $10 output per million tokens); Claude Haiku 5.5 still expected in coming weeks (no date announced) |
For what the lower price means for AI budgets, see our analysis of the Opus 5.5 and GPT-6 price cuts.
All claims in this section are Anthropic's; your results will depend on your tasks.
Anthropic says Opus 5.5 is "particularly good at long and sprawling jobs like codebase-wide migrations and audits." Per Anthropic's announcement, one early tester used it to audit and fix a 200,000-line codebase in under three hours; Opus 5 took more than 20 hours and used 2.5 times as many tokens on the same job.
Anthropic calls efficiency the area where Opus 5.5's advantage is "very clear": it costs less per token than Opus 5 and uses fewer tokens per task.
In an internal research test described in the announcement, 16 of 18 Opus 5.5 reports cleared Anthropic's quality bar, under which any invented figure or quote meant failure. Neither Fable 5.1 nor Opus 5 cleared it in any attempt. Anthropic also says Opus 5.5 puts the most important information first, uses less jargon, and follows the writing rules you give it.
Anthropic reports that Opus 5.5 matches or beats Opus 5 on prompt-injection resistance in every setting it tested, including coding, tool use, computer use and web browsing. That matters for agents that read content they didn't write.
Anthropic's own caution is worth repeating: at this level of capability, "benchmark margins have become a less reliable guide to real-world differences." Its comparison tables include OpenAI's latest models; we don't restate those vendor-reported head-to-heads as fact.
Anthropic's What's new in Claude Opus 5.5 page lists four breaking changes. The first three also apply to Fable 5.1, so teams that already moved to Fable have done some of this work.
| Change | Who's affected | What to do |
|---|---|---|
| Thinking can't be disabled | Code sending thinking: {"type": "disabled"} or a manual budget_tokens; both now return a 400 error |
Omit thinking or send {"type": "adaptive"}. Lower effort where you used to turn thinking off. |
| Forced tool use returns an error | Anyone using tool_choice any or tool, often to force structured JSON |
Keep tool_choice: auto and set strict: true (strict tool use), or move the schema to structured outputs. Say in the prompt when a tool applies. |
| Thinking blocks are tied to the model and conversation | Apps that switch models mid-conversation or edit earlier turns, system prompts or tools | Keep conversations append-only. Change instructions with mid-conversation system messages instead of edits. |
computer_20251124 tool rejected |
Computer-use agents on the Claude API or Google Cloud (Bedrock is unaffected) | Move to the computer_toolset_20260801 toolset and update your agent loop. |
Text between tool calls moves into thinking blocks |
Apps that stream progress notes to users; no error, the UI just goes quiet | Set thinking.display to return the text, and select content blocks by type, not position. |
| Preserved thinking enforced | API accounts created on or after August 31, 2026 (all cloud platforms included) | Edits before a thinking block return a 400 error; opt in to dropping blocks with the beta header, or avoid edits. |
Default effort is now medium |
Any request that omits effort (Opus 5 ran at high) |
Set effort explicitly and re-run your sweep. Leave room in max_tokens: Opus 5.5 tends to think more per turn at the same level. |
Model switching also has a direction. Per the docs, a conversation moving from Opus 5 onto Opus 5.5 keeps its reasoning, but one moving from Opus 5.5 to most other models continues without it; the request still succeeds. The Opus 5.5 migration guide shows before-and-after code for each change.
Anthropic says preserved thinking will expand to all users with upcoming model launches. If your integration edits earlier turns for compaction or injected reminders, plan that work now, even on an older account.
Claude.ai, Claude Code and Cowork handle thinking blocks for you, so app users have no code to change. They will notice these differences.
/model. Anthropic says fast mode (up to 2.5x speed at $8/$40 per million tokens) is available in Claude Code; on the API, fast mode is a research preview on the Claude API only.Availability inside third-party products such as Salesforce, Slack, HubSpot or Microsoft 365 follows each vendor's own schedule.
For compliance-sensitive teams in regulated industries, the public sector, healthcare or legal work, a model change belongs in change management. Record these points.
Opus 5.5 runs safety classifiers on every request. Per Anthropic's Help Center article on why Claude switches models, higher-risk offensive cybersecurity requests fall back to Opus 4.8, and dual-use biology and some frontier AI-development requests fall back to Opus 5. Attempts to extract the model's internal reasoning are blocked outright. The classifiers also check content from connectors, files, web search and memory, so a fallback can be triggered by text the user didn't type.
In the Claude apps, fallbacks show a notice and each response is labeled with the model that answered. On the API, a declined request returns HTTP 200 with stop_reason: "refusal" and a policy category; with server-side fallback (beta), the response's model field names the model that answered, per the refusals and fallback docs. Log that field and the stop reason. If you access Claude through a third-party tool, confirm how it reports fallbacks.
Anthropic says Opus 5.5 is available with zero data retention, like previous Opus models. One wrinkle: Anthropic's Cyber Verification Program doesn't include Opus 5.5 yet, and its support page says organizations on zero data retention aren't currently eligible for it. Opus 5.5 also carries Anthropic's text watermark for EU AI Act compliance. Anthropic says it adds no characters and carries no user or organization information.
Thinking can't be turned off, and on the API its content comes back empty by default. Decide whether your logs keep thinking content when you do request it. Requests to reproduce the model's reasoning verbatim can be refused under the reasoning_extraction category.
Anthropic reports Opus 5.5 scored best of any model to date on its automated behavioral audit of nearly 2,000 scenarios. In a containment test, it attempted to cross boundaries "around 85% less often" than Opus 5 or Claude Mythos 5.1. Anthropic is also candid about limits. It sees signs that the model "often suspects it is being evaluated," and says building evaluations that catch every failure before deployment "remains an unsolved problem." Record both in your review; the Opus 5.5 system card has the detail.
Add the model ID, platform, default effort, fallback models, data-retention setting and retirement commitment to your model inventory, and update the vendor-risk file.
Anthropic's models overview now recommends starting with Opus 5.5 for most workloads. It points to Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short in your evals.
| Model | Context | Max output | Price (input/output per MTok) | Latency | Default effort | Best fit |
|---|---|---|---|---|---|---|
| Claude Fable 5.1 | 1M | 128K | $10 / $50 | Slower | high |
Demanding reasoning and long-horizon agents |
| Claude Opus 5.5 | 1M | 128K | $4 / $20 | Moderate | medium |
Default for agentic coding and knowledge work |
| Claude Sonnet 5.5 | 1M | 128K | $2 / $10 | Fast | high on Claude Platform |
Well-scoped work, coding, and documents; benchmark against Opus for judgment-heavy tasks |
| Claude Haiku 4.5 | 200K | 64K | $1 / $5 | Fastest | Not supported | Simple, latency-sensitive tasks |
Sonnet 5.5 is available now at Sonnet 5’s $2 / $10 per-million-token standard API input/output prices; Anthropic positions it for well-scoped work and says Haiku 5.5 is still coming in the weeks ahead, with no date announced. Re-baseline your lower tiers against your own tasks. For Fable context, see our Claude Fable 5.1 guide.
Opus 5 is still active on Anthropic's model deprecations page, with retirement "not sooner than July 24, 2027" and no retirement date announced as of September 23, 2026. You don't have to move this week, but a week is enough to decide.
medium and at higher effort, and compare quality, latency and tokens per task against Opus 5.For broader adoption practices that still apply, see our Claude Opus 5 enterprise guide.
Vantage Point is an official Claude partner (Member, Claude Partner Network). Through our Claude implementation services, we help teams roll out Claude, build eval harnesses that make model upgrades routine, and connect Claude to Salesforce and HubSpot through MCP and connectors. Our compliance and security solutions cover the governance side: model inventories, fallback logging, and documentation for reviewers. Across 400+ engagements and 150+ clients, Vantage Point holds a 95% client retention rate and a 4.71/5.0 average engagement rating. Senior consultants only — no junior handoffs; the experts you meet are the experts who deliver.
A model upgrade is quicker when the evals, fallback logging and documentation are ready before you switch. Vantage Point can inventory where Claude runs in your organization, test Opus 5.5 on your real workflows, fix the breaking changes, and update your AI register for reviewers. Talk to Vantage Point about your Claude upgrade.
Claude Opus 5.5 is Anthropic's newest Opus model, released September 22, 2026, and the first in the Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry as claude-opus-5-5.
Anthropic lists four breaking changes: thinking can't be disabled, forced tool use returns an error, thinking blocks are tied to the model and conversation, and the older computer_20251124 tool is rejected on the Claude API and Google Cloud. Text between tool calls also moves into thinking blocks, so apps that stream progress updates can go quiet until they set a display option.
Anthropic's models overview recommends starting with Opus 5.5 for most workloads. It suggests Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Opus 5.5 at higher effort still fall short. Test both on your own tasks before deciding.
No. Adaptive thinking is always on for Opus 5.5, in the Claude apps and on the API. Requests that disable thinking or set a manual thinking budget return an error. Use the effort parameter to control thinking depth, latency and cost instead.
No. As of September 23, 2026, Anthropic's model deprecations page lists Opus 5 as active with retirement not sooner than July 24, 2027, and no retirement date has been announced. Anthropic gives at least 60 days' notice before retiring a publicly released model, so you can test Opus 5.5 first and move on your own schedule.
Yes. Anthropic says Opus 5.5 is available with zero data retention, like previous Opus models. Note that Anthropic's support page says organizations on zero data retention aren't currently eligible for its Cyber Verification Program, and that program doesn't include Opus 5.5 yet.
Opus 5.5 runs safety classifiers on every request. Per Anthropic, higher-risk offensive cybersecurity requests fall back to Opus 4.8, while routine secure coding stays on Opus 5.5. In the Claude apps, the response is labeled with the model that answered; on the API, check the stop reason and the model field, and log both.
Claude Sonnet 5.5 is available now at the same $2 input / $10 output per-million-token standard API list price as Sonnet 5. Anthropic says Claude Haiku 5.5 will follow in the coming weeks without an announced date. Re-run your evals for Sonnet 5 tasks now, and for Haiku workloads when Haiku 5.5 arrives.
Vantage Point is a boutique CRM consulting firm helping businesses transform with Salesforce, HubSpot, and AI — 150+ clients, 400+ engagements, and a 4.71/5 average engagement rating. Learn more at vantagepoint.io.