On September 22, 2026, Anthropic shipped Claude Opus 5.5 (claude-opus-5-5). It is a new flagship Opus-tier model with a 1M token context window, 128k max output tokens, and always-on adaptive thinking. List price is $4 per million input tokens and $20 per million output tokens. Cache hits are priced at $0.20 per million tokens, a 60% cut from the $0.50/MTok on Claude Opus 5. The model launched on the same day on the Claude API, Amazon Bedrock, Claude Platform on AWS (CCU billing), Google Cloud (Vertex AI), and Microsoft Foundry.
Three independent primary sources confirm the launch and the pricing: Anthropic's platform release notes (the dated September 22, 2026 entry), Anthropic's pricing page (a new row labelled "Claude Opus 5.5" sits above the Claude Opus 5 row, with cache reads footnoted at "0.05x the base input price"), and the "What's new in Claude Opus 5.5" reference page (which enumerates the breaking changes). Claude Code v2.1.280, released about three hours later at 16:38 UTC on the same day, confirms the headline claim verbatim: "Added Claude Opus 5.5 (claude-opus-5-5), now the default Opus model — 1M context, $4/$20 per Mtok with $0.20/Mtok cache reads."
Claude Code also flipped its default model in the same release: Pro and Team Standard plans now default to Opus (matching Max, Team Premium, and Enterprise, per the v2.1.280 changelog). That single change means a large slice of paying Claude Code users will be talking to Opus 5.5 the next time they hit Enter.
Three things changed that affect every agent stack running on Anthropic. One is a price drop. One is a context-window expansion. One is a default-model flip.
Price drop on the flagship Opus. Claude Opus 5 lists at $5/$25 per MTok. Claude Opus 5.5 lists at $4/$20. That is a 20% cut on input and a 20% cut on output. Cache reads move from $0.50/MTok to $0.20/MTok — a 60% cut. The cache multiplier is documented at the bottom of the pricing table: "Cache hits and refreshes on Claude Opus 5.5 are priced at 0.05x the base input price." For comparison, Claude Sonnet 5 also caches at $0.20/MTok (the 0.1x standard), and Claude Fable 5.1 / Mythos 5.1 cache at $0.25/MTok (the 0.025x custom tier). Opus 5.5 is now the cheapest cache-read tier Anthropic has ever offered, tied with Sonnet 5 at the cache-hit price floor.
1M context window default. Opus 5 was a 200k context model. Opus 5.5 is a 1M context model by default. Anthropic's "What's new in Claude Opus 5.5" page states the context window is 1M tokens out of the box, with 128k max output tokens. That puts Opus 5.5 at parity with Claude Fable 5.1 and Claude Mythos 5.1 on context, which shipped on September 1 at the $10/$50 price tier. The Opus tier is now both the cheaper tier and the larger-context tier — at least for input and cache reads; output is still 2.5x more expensive than Fable 5.1's $20/MTok output.
Default model flip in Claude Code. The v2.1.280 changelog says explicitly: "Changed the default model on Pro and Team Standard plans from Sonnet to Opus, matching Max, Team Premium, and Enterprise." Every Claude Code user on Pro or Team Standard who does not pin a model will get Opus 5.5 on the next session. If you are on Claude Code and your cost-per-turn suddenly looks different today, this is why.
There is one default behaviour change worth flagging: Anthropic's effort level handling. v2.1.280 also says "Changed an effort level saved before /effort became per-model to no longer apply to newly released models such as Opus 5.5; they start at their default until you pick a level." If you had a saved /effort setting, it does not carry over to Opus 5.5.
There are also four breaking API changes documented in "What's new in Claude Opus 5.5" that any production agent stack needs to know about:
1. Thinking cannot be disabled. On Opus 5 you could send thinking: {"type": "disabled"} at low effort. On Opus 5.5, a request that sends thinking: {"type": "disabled"} or thinking: {"type": "enabled", "budget_tokens": N} returns a 400. The error text is: "thinking.type.disabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior. Omit the thinking field, or send thinking: {"type": "adaptive"}. This is a code-level change in every Anthropic SDK that talks to Opus. 2. Forced tool use is not supported. tool_choice: {"type": "any"} and {"type": "tool", "name": "..."} both return 400 with the message tool_choice: type "tool" and "any" are not supported for this model. Use tool_choice: {"type": "auto"} (the default) plus strict: true with strict tool use, or move the schema to structured outputs. 3. Thinking blocks are tied to the model and the conversation. Opus 5.5 reads thinking blocks from Opus 5, Sonnet, and Haiku models, but not from Fable or Mythos models. Fable 5.1 and Mythos 5.1 read Opus 5.5 thinking blocks; no other model does. A conversation that moves from Opus 5.5 to a non-supported pair runs the turns after the switch without reasoning. There is also a prefix-binding check: as with Fable 5.1, accounts created on or after August 31, 2026 get this enforced by default; older accounts must opt in. Use the thinking-binding-controls-2026-08-01 beta header and set thinking.block_binding.prefix_mismatch_behavior to "drop_block" if you want to drop rather than error on a binding miss. 4. computer_20251124 is not supported on Claude API and Google Cloud. On Opus 5.5, a request that declares a computer_20251124 tool returns 400. Use computer_toolset_20260801. On Amazon Bedrock, the earlier tool keeps working — no change needed there.
There is also one additive capability: Fast mode (research preview) is available for Opus 5.5 on the Claude API only. Set speed: "fast" with the fast-mode-2026-02-01 beta header. It is not available on Bedrock, AWS Claude Platform, Google Cloud, or Microsoft Foundry. For everything else, Opus 5.5 supports per-message effort (beta), mid-conversation system messages, task budgets, prompt caching (with a 512-token minimum cacheable prompt), batch processing, the Files API, PDF support, vision, and server-side and client-side tools.
For a single agent on Opus-tier quality, Opus 5.5 is the cheapest Anthropic has ever priced the tier. If your stack runs a Sonnet or Haiku model today because Opus was too expensive at $5/$25, this is the day to re-run that math. At $4 input and $20 output, Opus 5.5's input is exactly the same price as Claude Sonnet 5 ($2/$10) multiplied by two — so Opus 5.5 is 2x the price of Sonnet 5 on input and 2x on output, but with the larger context window, the cache-read price parity, and the adaptive-thinking-by-default behaviour that Sonnet 5 does not have.
The cache-read math is the most underrated piece. At $0.20/MTok cache hits, a 1M-token cached request — which on Sonnet 5 / Opus 5.5 both fit comfortably in a single call — costs $0.20 to read back. That is the same cache price as Sonnet 5 and the cheapest cache price Anthropic has ever offered. For multi-turn agent loops that re-use system prompts and tool definitions, the marginal cost of context per turn just collapsed by a factor of 2.5x on Opus 5.5 vs. Opus 5. If your cost model assumes a per-turn cache-read rate above $0.20/MTok, that assumption is now stale.
The default-model flip in Claude Code matters even if you do not use Opus yourself. If your team uses Claude Code on Pro or Team Standard and has no pinned model, every developer is now spending at the Opus rate on their next session. If the team was previously optimising around Sonnet pricing, cost reports will jump. If you have not pinned a model in .claude/settings.json or via --model flag, pin one.
The breaking API changes are the operational story. Three of the four — thinking cannot be disabled, forced tool use returns 400, and the computer_20251124 rejection — are request-time errors. Code paths that worked against Opus 5 will return 400 against Opus 5.5 if you did any of: (a) sent thinking.type: "disabled", (b) forced a tool call with tool_choice: "any" or "tool", (c) attached a computer_20251124 tool to a request that hits the Claude API or Google Cloud. If you are migrating an agent fleet from Opus 5 to Opus 5.5, these are your migration test cases.
The thinking-block-binding check is the subtler one. Opus 5.5 reads thinking blocks from Opus 5, Sonnet, and Haiku — but not from Fable or Mythos. If you have a multi-model agent that hops between Opus and Fable on different turns, that pattern still works on Opus 5.5. If you have a pattern that hops between Opus and Mythos, the thinking context does not carry. The pricing page footnote that calls out the new 0.05x cache multiplier is a hint: this is part of the same generation's behaviour as Fable 5.1, not a regression.
All claims above are documentation comparisons. No firsthand production-harness test was executed for this article. Sources, in the order they appear above:
Independent corroboration: the Anthropic pricing page is the same source as the release notes, but the table structure is independently verifiable — the row order is Fable 5.1 / Mythos 5.1 / Fable 5 / Mythos 5 / Opus 5.5 / Opus 5 / Opus 4.8 / Opus 4.7 / Opus 4.6 / Opus 4.5 / Opus 4.1 / Opus 4 / Sonnet 5 / Sonnet 4.6 / Sonnet 4.5 / Sonnet 4 / Haiku 4.5 / Haiku 3.5. Opus 5.5 sits in the spot a release notes claim would predict — between Mythos 5 and Opus 5 — and has its own footnote superscript on the cache-hit cell.
What we did not independently run for this article:
claude-opus-5-5 was placed. The "thinking cannot be disabled" 400 behaviour, the tool_choice 400 behaviour, and the computer_20251124 400 behaviour are reported from Anthropic's own documentation, not from a live request.Cost. For a representative long-running agent task that reads ~500k cached input tokens and writes ~20k output tokens per turn, on Opus 5.5 the marginal cost is roughly: cache reads = 500,000 × $0.20 / 1,000,000 = $0.10; output = 20,000 × $20 / 1,000,000 = $0.40. Total $0.50 per turn before any non-cached input. That compares to Opus 5's same workload at roughly $0.25 cache + $0.50 output = $0.75 per turn, a 33% saving. These are list prices and exclude prompt-cache writes, tool-result tokens, and any non-cached input tokens. Real workloads will be higher; treat the $0.50 figure as a lower bound for that workload shape.
The Anthropic pricing page footnote about regional pricing still applies on Opus 5.5: Bedrock regional endpoints and Google Cloud regional/multi-region endpoints carry a 10% premium over global endpoints, and the Claude API is global by default. For first-party data residency pricing on Opus 5.5, see the data residency section of the same page.
Risk. Three concrete risks for an existing Opus 5 production stack:
thinking: {"type": "disabled"} or uses tool_choice: "any" / "tool" will return 400 on first request. Test your agent fleet's request shapes before flipping the model ID.computer_20251124 returns 400. Migrate to computer_toolset_20260801 before flipping the model ID. Amazon Bedrock is unaffected.A subtler risk is the default-model flip in Claude Code itself. If you have team members on Pro or Team Standard who are not pinning models, they will start spending at Opus 5.5 rates. This is a real-dollar cost change for teams that have not budgeted for it. Pin models in .claude/settings.json or via --model if you do not want a default flip.
Limitations. Fast mode is research preview, and only on the Claude API — not on Bedrock, AWS Claude Platform, Google Cloud, or Microsoft Foundry. If your stack runs on a non-API platform and you were planning to use Fast mode, that path is not open yet. Token inflation is inherited from the 4.7+ tokenizer family; if you are porting prompts from Sonnet 4.6 or earlier, expect ~30% more tokens for the same text. And like Fable 5.1 and Mythos 5.1, Opus 5.5 requires 30-day data retention by default; it is not available under zero data retention unless expressly authorized.
Anthropic shipped the right price for the right tier. $4/$20 with 1M context and the cheapest cache reads Anthropic has ever offered is a deliberate move to make Opus-tier quality the default tier for serious agent work — not a niche tier for the few who could afford it. The default-model flip in Claude Code says the same thing in product language: the company thinks Opus 5.5 should be the model you use by default, and they are willing to flip a paid plan's default to make that happen.
The breaking changes are real but manageable. Three are request-level errors you can catch in staging: thinking-disabled, forced tool use, computer_20251124 on Claude API or Google Cloud. The fourth — thinking-block binding tied to the model and conversation — is a behaviour change, not an error, and is the one worth understanding before you migrate. For most stacks that already use adaptive thinking and do not force tool calls, Opus 5.5 is a drop-in: cheaper input, cheaper output, cheaper cache, larger context, same SDK.
For new agent builds, Opus 5.5 is now the right default starting point for the Opus tier. Sonnet 5 at $2/$10 is still cheaper for high-volume, low-context workloads. Fable 5.1 at $10/$50 is still the top tier for the most expensive reasoning jobs where $50/MTok output is acceptable. Opus 5.5 sits in the middle: 2x Sonnet 5's price for context parity and cache-read parity, with adaptive thinking by default.
The headline risk is the cost surprise for teams who did not pin a Claude Code default. The headline opportunity is that anyone paying Opus-tier prices today should re-run their cost model with the new numbers.
Today
1. Pin your Claude Code default model in .claude/settings.json or with --model, explicitly. If you do not want to switch every Pro or Team Standard user to Opus 5.5 by accident, do this before they hit Enter on a new session. The default is now Opus 5.5. 2. Audit your request shapes for the three request-level errors: thinking.type: "disabled", tool_choice: "any" or "tool", and computer_20251124 tools attached on Claude API or Google Cloud. Each one returns a 400 on Opus 5.5. 3. Update cost models. Replace Opus 5 rates ($5/$25 input/output, $0.50/MTok cache) with Opus 5.5 rates ($4/$20, $0.20/MTok) in any forecast. For workloads with high cache-hit rates, the savings are largest. 4. Rerun cache-hit assumptions. Opus 5.5 cache reads are the cheapest Anthropic has ever offered, tied with Sonnet 5 at $0.20/MTok. If your cost model assumed $0.50/MTok on the previous Opus tier, you are now 2.5x over-provisioning.
This week
5. Migrate the computer-use path if you are on Claude API or Google Cloud. Move from computer_20251124 to computer_toolset_20260801 before flipping any agent to Opus 5.5. Amazon Bedrock is unchanged. 6. Test cross-model thinking-block handoffs. If your agent workflow hops between Opus 5.5 and an older model, validate that the handoff behaves as expected. The pricing page documents which models can read each other's thinking blocks; the "What's new in Opus 5.5" page is more specific. 7. Decide on Fast mode. If you are running on the Claude API and have latency-sensitive paths, Fast mode is available as a research preview with the fast-mode-2026-02-01 beta header. It is not on Bedrock, AWS Claude Platform, Google Cloud, or Microsoft Foundry.
Skip if not in scope
claude-opus-5-5 and inherit the price drop. Your platform-specific behaviour is unchanged from Opus 5 (except the regional endpoint 10% premium, which still applies). The computer_20251124 rejection does not apply on Bedrock.