Is Claude Opus 5.5 better than Opus 5?

Yes, on every benchmark Anthropic published on September 22, 2026: Terminal-Bench 4.0 rose from 52.3% to 66.4%, FrontierCode v1.1 from 48.0% to 54.4%, CursorBench 4.0 from 46.6% to 57.8%, GDPval-AA v2.1 from 1708 to 1846 Elo, OSWorld 2.0 from 74.0% to 81.8%, and Humanity's Last Exam from 63.6% to 67.7%. Artificial Analysis independently scored Opus 5.5 at 58 on its Intelligence Index versus 51 for Opus 5. It also writes more clearly, generates output over 30% faster, and scores better on Anthropic's alignment audit and on prompt-injection resistance.

Is Claude Opus 5.5 cheaper than Opus 5?

Per token, yes: $4/$20 per million input/output versus $5/$25 (20% less), cache reads $0.20 versus $0.50 (60% less), and 5-minute cache writes $5 versus $6.25. Anthropic measures a 40% lower cost on typical workloads at default settings because Opus 5.5 also uses fewer tokens per task. One caveat from Artificial Analysis: at max effort Opus 5.5 emits about 119,000 output tokens per Intelligence Index task against 73,000 for Opus 5, so cost per task at max is roughly level; the savings appear at medium, high and xhigh effort.

What breaks when I switch from Opus 5 to Opus 5.5?

Four API changes return 400 errors: disabling thinking or setting a manual budget_tokens (thinking is always adaptive; use effort instead), forcing tool use with tool_choice 'any' or a named tool, replaying thinking blocks after the system prompt or tools changed (enforced by default for accounts created on or after August 31, 2026), and using the legacy computer_20251124 tool on the Claude API or Google Cloud. Text between tool calls also moves into thinking blocks that are empty at the default display setting. Code that ran Opus 5 with adaptive thinking on and tool_choice auto needs no change.

Should I stay on Claude Opus 5?

Stay temporarily if you depend on forced tool use, disabled thinking, or the old computer-use tool and cannot refactor this week, or if you run offensive-security or biology workloads that Opus 5.5's Fable-class safeguards would re-route to Opus 4.8 or Opus 5 anyway. Otherwise switch: Opus 5.5 is cheaper, faster, and stronger, and Anthropic has positioned it as the new default Opus. No deprecation date for Opus 5 has been announced as of September 23, 2026.

Quick Answer

Claude Opus 5.5 vs Opus 5: Should You Upgrade? (Sep 2026)

Published: September 23, 2026

The short answer

Upgrade. Claude Opus 5.5, released September 22, 2026, beats Claude Opus 5 on every published benchmark, costs 20% less per token and 60% less per cached token, writes more clearly and generates output over 30% faster. The only reasons to wait are four API breaking changes and Fable-class safeguards that re-route some security and biology work. Plan the migration around those, then move.

Side by side

	Claude Opus 5.5	Claude Opus 5
Released	September 22, 2026	July 24, 2026
Price (in / out per MTok)	$4 / $20	$5 / $25
Cache read	$0.20 (5% of input)	$0.50 (10% of input)
5m / 1h cache write	$5 / $8	$6.25 / $10
Fast mode	$8 / $40, up to 2.5x speed	$10 / $50
Context / max output	1M / 128K	1M / 128K
Knowledge cutoff	June 2026	Earlier 2026
Thinking	Adaptive, always on; default effort medium	Adaptive on by default; can be disabled at high or below
Forced tool use	Not supported	Supported
Computer use (API, GCP)	`computer_toolset_20260801` only	Toolset or legacy `computer_20251124`
AA Intelligence Index (max)	58	51
Output tokens per AA task (max)	~119K	~73K
Safeguards	Fable 5.1-class (cyber, bio, distillation)	Standard Opus

Benchmarks: every row moved

Benchmark	Opus 5.5	Opus 5	Gain
Terminal-Bench 4.0	66.4%	52.3%	+14.1 pts
FrontierCode v1.1 (Main)	54.4%	48.0%	+6.4 pts
CursorBench 4.0	57.8%	46.6%	+11.2 pts
GDPval-AA v2.1 (Elo)	1846	1708	+138
AutomationBench (Zapier)	40.0%	26.9%	+13.1 pts
Humanity’s Last Exam (tools)	67.7%	63.6%	+4.1 pts
Terminal-Bench-Science 0.1	58.7%	29.0%	+29.7 pts
OSWorld 2.0 (partial)	81.8%	74.0%	+7.8 pts
Chartography (tools)	89.0%	83.4%	+5.6 pts

Terminal-Bench-Science doubling is the standout: Anthropic’s science-agent harness went from below GPT-5.6 Sol to within six points of GPT-6 Astra. Artificial Analysis’s independent Terminal-Bench 4.0 run put Opus 5.5 at 59.6% (+11 over Opus 5), a smaller but consistent gain.

The cost story has two halves

Per-token prices fell. Opus pricing had been frozen at $5/$25 since Opus 4.5 in November 2025. Opus 5.5 is $4/$20, and the cache-read price fell from $0.50 to $0.20. For an agentic session where 90% of input is cached, blended input cost per million tokens drops from about $0.95 to about $0.58, a 39% cut before counting output.

Tokens per task depend on effort. Anthropic’s 40% saving applies at default (medium) effort. Artificial Analysis found that at max effort Opus 5.5 thinks 1.6x longer than Opus 5 at max, which cancels the price cut. Simon Willison hit the 128K output ceiling twice on max with an SVG prompt, each attempt costing $2.56 and taking nearly 20 minutes. Practical rule: migrate at medium or high and only raise effort where you measure a gain.

Worked example, 1,000 long agentic tasks with 300K input (90% cached) and 30K output each:

Opus 5: input ≈ $285, output $750 → ≈ $1,035
Opus 5.5 (same tokens): input ≈ $174, output $600 → ≈ $774 (25% less), and Anthropic’s tests suggest fewer tokens on top.

What actually changed under the hood

Anthropic describes Opus 5.5 as needing less compute to serve than Opus 5, which is what funds the price cut. Observable changes: faster output, clearer writing that leads with the conclusion, better instruction-following on style rules, the best scores yet on Anthropic’s ~2,000-scenario behavioral audit, an 85% reduction in containment-boundary attempts versus Opus 5, and prompt-injection resistance that ties Fable 5.1 as the lowest success rate on Gray Swan’s benchmark. GitHub reported it used “among the fewest tokens and steps we measured” in Copilot CLI and VS Code, solving more terminal tasks than Opus 5 in less than half the steps.

Upgrade decision by workload

Workload	Recommendation	Why
Long agentic coding, migrations, audits	Upgrade now	Terminal-Bench +14 pts; 200K-line audit in <3h vs 20h+
Knowledge work, reports, financial models	Upgrade now	GDPval +138 Elo; 16/18 fact-checked reports passed vs 0 for Opus 5
Code review at low effort	Upgrade now	Deloitte: 72% of known bugs caught at low effort vs 56% for Opus 5 at high
Computer-use agents	Upgrade, refactor first	OSWorld +7.8 pts, but legacy `computer_20251124` rejected on API/GCP
Pipelines using forced tool use	Upgrade after refactor	`tool_choice: any/tool` returns 400; use strict tools or structured outputs
Offensive security, wet-lab biology	Stay or apply for verification	Fable-class safeguards route these to Opus 4.8 / Opus 5
Max-effort batch jobs	Test before switching	1.6x output tokens at max can erase the savings

Migration checklist

Change the model ID to claude-opus-5-5 in one non-critical route.
Delete any thinking: {"type": "disabled"} or budget_tokens; set output_config.effort instead.
Replace tool_choice: any/tool with auto plus strict: true, or structured outputs.
Swap computer_20251124 for {"type": "computer_toolset_20260801"} and drop the beta header.
Keep conversations append-only; use mid-conversation system messages instead of editing history so thinking-block binding never trips.
Select content blocks by type, not position, and set display if you stream inter-tool text to users.
Re-baseline token budgets at medium effort; compare against Opus 5 for a week before raising effort.

Full guide: How to migrate to Claude Opus 5.5. For the wider picture, see What is Claude Opus 5.5? and Opus 5.5 vs Fable 5.1 vs GPT-6 Astra.

Last verified: September 23, 2026.