Claude Opus 5.5 Is the New Default — Here Is What Changes

Anthropic shipped Claude Opus 5.5 on September 22 and immediately made it the default model across Claude Code, the Claude app, and Cowork for Pro, Max, and Team subscribers. It outperforms Fable 5.1 on every benchmark Anthropic published at launch, costs roughly 40% less per task than Opus 5, and is now the model running every time you open Claude Code. There are also four breaking API changes that will either silently misbehave or throw hard errors in existing agent integrations — the upgrade is not free.

The Economics Are the Story

The headline numbers look clean: input tokens drop 20% to $4 per million, output tokens drop 20% to $20 per million. But the bigger savings sit in cache reads, which fall 60% — from $0.50 to $0.20 per million tokens. For agentic coding workflows where stable system prompts, repo context, and tool definitions get re-read on every turn, cache reads represent the majority of actual spend. That 60% reduction compounds hard.

Anthropic reports roughly 40% total cost reduction per typical workload, combining lower list pricing with fewer tokens consumed per task. Early testers reported completing a 680,000-line code migration in under a day. A 200,000-line audit-and-fix that took Opus 5 over 20 hours finished in under 3 hours. One HAProxy C-to-Rust translation ran 9.5 hours and cost 51% less than Fable 5.1 on the same task.

One caveat worth flagging: at maximum effort, Opus 5.5 generates roughly 4x the output tokens of GPT-6 Astra on benchmark tasks, and the cost advantage nearly disappears. Anthropic’s 40% savings claim holds at medium effort — which is also the new default. Run the model at max effort on heavy reasoning workloads and recalculate before committing.

The Benchmark Picture

Coding benchmarks show a genuine step up. Opus 5.5 hits 89.9% on SWE-bench Pro (Fable 5.1: 81.2%, Opus 5: 79.2%), 66.4% on Terminal-Bench 4.0 (Opus 5: 52.3%), and 57.8% on CursorBench 4.0 at max effort. It ranks first on the Artificial Analysis Intelligence Index, five points ahead of both GPT-6 Astra and Fable 5.1. At medium effort, CursorBench scores 52.5% — already above Fable 5.1’s maximum-effort score of 51.8%, at roughly 80% lower cost per task.

Benchmark	Opus 5.5	Fable 5.1	Opus 5
SWE-bench Pro	89.9%	81.2%	79.2%
Terminal-Bench 4.0	66.4%	55.8%	52.3%
CursorBench 4.0	57.8%	51.8%	46.6%
Humanity’s Last Exam	67.7%	65.6%	63.6%

Anthropic itself acknowledged that “benchmark margins have become a less reliable guide to real-world differences.” That is unusually candid, and accurate. The gap between Opus 5.5 and Fable 5.1 is narrower in practice than headline scores suggest.

Four Breaking Changes — Fix These Before You Migrate

Opus 5.5 ships with mandatory adaptive thinking and a reworked tool model. Four patterns that worked on Opus 5 will fail on 5.5. Check the full migration guide for code examples, but here is the summary:

• Thinking cannot be disabled. Sending thinking: {"type": "disabled"} returns a 400 error. Replace it with output_config: {"effort": "low"} to minimize latency and cost.
• Forced tool choice is gone. tool_choice: {"type": "any"} and {"type": "tool", "name": "..."} both return 400. Switch to {"type": "auto"} and add strict: true to your tool definitions. Verify the tool was actually called and retry if not.
• Thinking blocks are bound to the conversation. Any change to the system prompt, tool list, or message history mid-session invalidates thinking blocks — enforced for accounts created after August 31, 2026. Keep message history append-only, declare all tools at session start. If your agent trims history, enable drop_block behavior.
• Computer use tool is deprecated on Claude API and Vertex AI. The computer_20251124 tool returns 400. Migrate to computer_toolset_20260801. Amazon Bedrock still accepts the old tool. The agent loop also changes: actions now arrive as individual tool_use blocks with the action name in block.name, and results must include "toolset_name": "computer".

Community Verdict: Cautiously Positive

Developer reaction on Reddit and Hacker News skewed positive but with friction. A practical workflow split is emerging: “plan with Fable, implement with Opus” — Fable 5.1 for orchestration and broad research, Opus 5.5 for execution and coding. Experienced users point out that Fable still holds an edge for wide-domain reasoning spanning many subjects simultaneously.

The pacing controversy is worth a moment. Anthropic publicly called for deliberate pacing at the frontier. Releasing a meaningful capability upgrade at 40% lower cost 21 days after Fable 5.1 raised eyebrows on Hacker News within minutes of the announcement. The company is free to update its position — competitive pressure from OpenAI’s same-day GPT-6 Sol and Luna releases is real — but the gap between stated philosophy and shipping cadence was noticed.

What to Do Now

If you are running Opus 5 today, migrate. The economics favor it and the coding performance is a clear improvement. Run at medium effort by default — it already beats Fable 5.1’s max-effort scores at a fraction of the cost. Address the four breaking changes before deploying to production, particularly the thinking-block conversation binding if your agent modifies history mid-session. If you use computer use on Claude API or Vertex AI, computer_toolset_20260801 is non-negotiable.

If you are on Fable 5.1 and primarily use it for planning, orchestration, or broad research, the calculus is less clear. Fable retains an edge there. The “plan with Fable, implement with Opus” split is a reasonable starting point. Track cost per completed task — not cost per token — and let the numbers decide. See Vellum’s benchmark breakdown for a detailed effort-level cost comparison before committing either way.