TL;DR: Claude Opus 5.5 (released September 22, 2026) replaces Opus 5 at $4/$20 per million tokens — 20% less than Opus 5’s $5/$25 — with cache reads cut 60% to $0.20/M. On Anthropic’s own launch table it beats Claude Fable 5.1, the $10/$50 flagship, on every listed coding benchmark. If you use Claude Code v2.1.280+, you are probably already on it: default and opus now resolve to Opus 5.5 on nearly every plan and provider. Cursor has it in the model picker. Cline does not have it in the built-in catalog as of v4.1.20 — and Cline’s Anthropic provider default still silently resolves to Fable 5.1 at 2.5x the price.

	Claude Opus 5.5	Claude Opus 5	Claude Sonnet 5	Claude Fable 5.1
Input / output per MTok	$4 / $20	$5 / $25	$2 / $10	$10 / $50
Cache read per MTok	$0.20 (0.05x)	$0.50 (0.1x)	$0.20 (0.1x)	$0.25 (0.025x)
Terminal-Bench 4.0 (vendor-run)	66.4%	52.3%	—	55.8%
Default effort	medium	high	high	high
Heavy cached session (math below)	~$3.80	~$6.25	~$2.50	~$8.00
The catch	Effort default dropped a notch; 4 breaking API changes	Superseded — same surface, worse price	Widest quality gap on the hardest 20% of work	Loses to Opus 5.5 on Anthropic’s own table at 2.5x the price

Honest take: Opus 5.5 is the rare model launch where the right move is obvious. It is cheaper than the model it replaces, faster (30% higher output speed, per Anthropic), and ahead of both Opus 5 and Fable 5.1 on the vendor’s own agentic-coding benchmarks. Switch — unless you run a custom BYOK harness that touches one of the four breaking changes below, or you’re on a budget where Sonnet 5 at half the price was already good enough. The one thing to do today even if you change nothing else: pin your effort level, because Opus 5.5 quietly defaults to medium where every other current Claude model defaults to high.

What is Claude Opus 5.5 and what changed from Opus 5?

Claude Opus 5.5 is Anthropic’s new default flagship for agentic coding, released September 22, 2026. The API model ID is claude-opus-5-5 (anthropic.claude-opus-5-5 on Amazon Bedrock). Specs verified against Anthropic’s model docs on September 23, 2026: 1M-token context window, 128K max output (300K on the Batch API with the output-300k-2026-03-24 beta header), knowledge cutoff June 2026, retirement no sooner than September 22, 2027. It is live on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, plus claude.ai Pro/Max/Team/Enterprise plans and Claude Code.

Four things changed that matter for coding use:

1. Every price went down. Input $5 → $4, output $25 → $20, 5-minute cache writes $6.25 → $5, cache reads $0.50 → $0.20. Anthropic also claims the model finishes tasks with fewer tokens, putting typical workloads around 40% cheaper than Opus 5 in total — the per-token cut alone accounts for most of that in cache-heavy sessions (verified math below).
2. Output is about 30% faster than Opus 5, per Anthropic’s launch material. Comparative latency is listed as “Moderate” — a tier faster than Fable 5.1’s “Slower.”
3. The default effort dropped from high to medium. No other current Claude model does this. Details below, because it skews every casual before/after comparison.
4. Four breaking API changes for anyone calling Opus 5 directly through a custom harness — covered in the migration section.

How much does Claude Opus 5.5 cost per coding session?

About $1.30 for a heavy uncached agentic session, and about $3.80 for a cache-heavy marathon session — versus $1.63 and $6.25 on Opus 5. All prices verified against Anthropic’s official pricing docs on September 23, 2026:

Per million tokens	Opus 5.5	Opus 5	Sonnet 5	Fable 5.1
Input	$4.00	$5.00	$2.00	$10.00
Output	$20.00	$25.00	$10.00	$50.00
Cache read	$0.20	$0.50	$0.20	$0.25
5m cache write	$5.00	$6.25	$2.50	$12.50
Batch API	50% off	50% off	50% off	50% off
Fast mode	$8 / $40	$10 / $50	—	—

For the uncached 200K-input / 25K-output session shape we’ve used since the Opus 5 default cost analysis: Opus 5.5 runs $0.80 input + $0.50 output = $1.30, versus $1.63 on Opus 5, $0.65 on Sonnet 5, and $3.25 on Fable 5.1. At 5 sessions a day, 22 working days: $143/month versus $179 on Opus 5.

The bigger cut is in cached agentic loops — which is what Claude Code, Cursor agents, and Cline actually run. Take the marathon-session shape from the Fable 5.1 analysis: 6M tokens of cache reads, 200K of 5-minute cache writes, 100K fresh input, 60K output:

Cost component	Opus 5.5	Opus 5	Sonnet 5	Fable 5.1
6M cache reads	$1.20	$3.00	$1.20	$1.50
200K cache writes	$1.00	$1.25	$0.50	$2.50
100K fresh input	$0.40	$0.50	$0.20	$1.00
60K output	$1.20	$1.50	$0.60	$3.00
Session total	$3.80	$6.25	$2.50	$8.00

That is 39% below Opus 5 on identical token counts — pure price cut, before Anthropic’s “finishes tasks with fewer tokens” claim adds anything. If the token-efficiency claim holds in practice, real sessions come in under this table, not over it. Two such sessions per weekday drops the monthly bill from about $275 (Opus 5) to about $167.

Notice the cache-read column: Opus 5.5 now matches Sonnet 5’s $0.20/M exactly. In read-dominated sessions, the Opus-over-Sonnet premium shrinks from 2.5x to about 1.5x ($3.80 vs $2.50 above). Anthropic keeps pricing its coding models to be run for hours.

Is Opus 5.5 actually better at coding than Opus 5 and Fable 5.1?

On Anthropic’s own launch table, yes — it beats both on every listed benchmark, including the $10/$50 Fable 5.1. All scores below are vendor-run, published September 22, 2026, with no independent verification yet:

Benchmark	Opus 5.5	Fable 5.1	Opus 5	GPT-6 Astra	GPT-5.6 Sol
Terminal-Bench 4.0	66.4%	55.8%	52.3%	57.9%	37.3%
FrontierCode v1.1	54.4%	50.3%	48.0%	53.3%	47.5%
CursorBench 4.0	57.8%	51.8%	46.6%	—	41.7%
GDPval-AA v2.1	1846	1735	1708	1542	1588
AutomationBench	40.0%	31.4%	26.9%	41.4%	28.8%
OSWorld 2.0	81.8%	80.7%	74.0%	—	—

Three caveats before you repeat these numbers:

• The headline Terminal-Bench score ran at xhigh effort (standard error ±2.6 points, per Vellum’s benchmark breakdown), while GPT-6 Astra’s 57.9% ran at high. Different effort settings across vendors make cross-model rows softer than same-vendor rows. The Opus 5.5 vs Opus 5 vs Fable 5.1 comparison is the reliable part of this table.
• GPT-6 Astra still wins AutomationBench (41.4% vs 40.0%). Anthropic published the loss, which is to its credit, but “beats GPT-6 Astra everywhere” — a claim already circulating — is not what the table says.
• Some secondary coverage attaches an “89.9% SWE-bench Pro” score to this launch. That number is not in Anthropic’s published table, which uses Terminal-Bench 4.0, FrontierCode, and CursorBench instead. Don’t cite it until it appears somewhere first-party or on an independent leaderboard.

The Fable 5.1 result is the strategic story. Fable 5.1 launched September 1 at $10/$50 as the top “Mythos-class” model; three weeks later Anthropic’s own table shows the $4/$20 Opus beating it on every agentic-coding row. What’s left for Fable 5.1 is its 0.025x cache reads ($0.25/M — cheaper relative to its base price but now more expensive in absolute terms than Opus 5.5’s $0.20/M) and whatever long-horizon behavior didn’t make the benchmark table. If you bought Fable 5.1 for coding, this launch is your exit.

The medium effort default: the quiet change that skews comparisons

Claude Opus 5.5 defaults to medium effort; Opus 5, Sonnet 5, and Fable 5.1 all default to high. Verified against Anthropic’s model comparison table on September 23, 2026 — Opus 5.5 is the only current model with a lowered default.

This has two practical consequences:

1. Naive A/B tests understate the new model. Send the same request at defaults to Opus 5 and Opus 5.5 and you’re comparing high against medium. The 66.4% Terminal-Bench headline was measured at xhigh — two notches above what you get out of the box.
2. Default-setting sessions are cheaper but shallower. Lower effort means fewer thinking tokens and more consolidated tool calls. Anthropic’s 40%-cheaper claim is partly this — some of the saving comes from the model doing less deliberation unless you ask for it.

If you’re benchmarking or migrating a harness, pin effort explicitly (output_config: {"effort": "high"} on the API) and compare like for like. If you just want cheap daily coding, the medium default is arguably the right call — but it should be your call, not a silent one.

Claude Code: you are probably already running Opus 5.5

Claude Code switched its default alias to Opus 5.5 on launch day. Verified against the official model-config docs on September 23, 2026: default resolves to Opus 5.5 on Pro, Max, Team, Enterprise, and Anthropic API billing, and on Claude Platform on AWS, Amazon Bedrock, and Google Cloud’s Agent Platform (Microsoft Foundry is the outlier, defaulting to Sonnet 4.5). The opus alias also now means Opus 5.5. The requirement: Claude Code v2.1.280 or later — older versions still resolve opus to Opus 5.

Check and pin your model in one command:

/model            # opens the picker — Enter saves as default, "s" applies to this session only
/model opus       # latest Opus (5.5 on v2.1.280+)
/model opusplan   # Opus for plan mode, Sonnet 5 for execution — the budget hybrid

For flat-rate subscribers this is strictly good news: same monthly price, better model, and because Opus 5.5 burns fewer tokens per task (per Anthropic), your usage limits stretch further. For API-billed Claude Code the math is the session table above — roughly 20-39% off, depending on cache mix.

One migration note from the July 2026 default flip still applies: when Claude Code changed defaults to Opus 5, API-billed users who never chose a model saw their per-session cost jump without any settings change. This flip goes the other direction — costs drop — but the lesson stands: if per-token cost matters to you, pin a model instead of riding default.

How do you use Claude Opus 5.5 in Cursor?

Cursor listed Claude Opus 5.5 in the model picker on launch day — select it under Settings → Models, no BYOK required. Per Cursor’s model docs (verified September 23, 2026 via search cross-reference; cursor.com blocks our direct fetch): it bills at the same $4/$20 API rates through Cursor’s “Other Models” usage pool, cache reads at $0.20/M, and it set a new CursorBench high score, ahead of Fable 5.1. On legacy request-based plans it requires Max Mode. A fast variant (claude-opus-5-5-fast) is available at $8/$40 — the same 2x multiplier Anthropic charges for fast mode on the API, still a research preview there.

For Cursor Pro ($20/month) users the calculus is unchanged from Opus 5: included usage covers moderate Opus work, and heavy agentic users burn through the pool faster than a flat-rate Claude Code plan would allow. What did change is the BYOK break-even — at $4/$20 with $0.20 cache reads, an API key routed through Cursor now costs about 39% less per marathon session than the same workflow ten days ago. GitHub Copilot also added Opus 5.5 to its model picker on September 22 for Pro/Pro+ and (admin-enabled) Business/Enterprise plans, so every major harness except one now carries it. That one:

Cline can’t select Opus 5.5 yet — and the Fable 5.1 default trap is still live

As of Cline v4.1.20 (the latest release as of September 23, 2026), claude-opus-5-5 is not in Cline’s built-in model catalog. We verified this in the source: the v4.1.20 changelog’s catalog refresh (209 providers, 6,237 models) predates the launch and contains no Opus 5.5 entry, and Cline’s Anthropic provider only accepts models from that curated catalog — free-text model IDs are reserved for the openai-compatible, ollama, lmstudio, litellm, and vertex providers. Until a catalog refresh ships, you cannot type the new ID into the Anthropic provider and go.

Two working routes in the meantime, both verified against Cline’s source on September 23, 2026:

1. OpenRouter provider. Cline fetches OpenRouter’s model list live every time the extension panel loads, and OpenRouter listed anthropic/claude-opus-5.5 at launch. Add an OpenRouter key, refresh, select it. You pay OpenRouter’s routing margin, but you’re on the new model today.
2. Vertex provider with a typed ID. Vertex is one of the providers that accepts custom model IDs — Google Cloud users can enter claude-opus-5-5 directly.

The sharper problem is what Cline runs when you don’t choose. Since v4.1.17, Cline’s Anthropic provider default resolves to Claude Fable 5.1 — the $10/$50 model that Opus 5.5 just beat across Anthropic’s entire launch table. We flagged this as a cost trap when Fable 5.1 shipped; it’s now a cost trap with no quality argument left. If you run Cline against an Anthropic key and have never pinned a model, you are paying 2.5x Opus 5.5’s rate for a model that scores lower on the vendor’s own coding benchmarks. Open your provider settings and pin something — today that means Opus 5 ($5/$25, in the catalog) or Sonnet 5 ($2/$10), then switch to Opus 5.5 when the catalog update lands.

Four breaking changes if you call the API directly

BYOK harnesses moving from claude-opus-5 to claude-opus-5-5 hit four breaking changes, verified against Anthropic’s migration docs on September 23, 2026:

1. Thinking can’t be disabled. thinking: {"type": "disabled"} returns a 400 at every effort level (Opus 5 accepted it at high and below). Omit the parameter and control depth with effort instead — low is the new “as little thinking as possible.”
2. Forced tool use is gone. tool_choice: {"type": "any"} and {"type": "tool", ...} return a 400. Use auto plus an explicit prompt instruction, strict: true on the tool for schema-valid arguments, or structured outputs if the forced call only existed to get JSON back.
3. Thinking blocks are tied to the model and the conversation (“preserved thinking”). Editing earlier turns invalidates them; API accounts created on or after August 31, 2026 get a hard 400 on edited history. Make your harness append-only.
4. The older computer_20251124 computer-use tool is rejected on the Claude API and Google Cloud — only computer_toolset_20260801 works.

Plus one change that breaks nothing but confuses UIs: text the model emits between tool calls now arrives as thinking blocks whose text is empty at the default display setting. If your harness streams that text to users as progress updates, it goes silent mid-loop until you set display to a value that returns it. The first three breaking changes match Fable 5.1’s behavior, so harnesses already updated for Fable 5.1 mostly carry over.

When should you NOT switch to Opus 5.5?

Four cases, honestly:

• Your harness depends on forced tool_choice, disabled thinking, or computer_20251124. The 400s above are hard errors, not warnings. Budget a migration pass first; Opus 5 remains available in the meantime.
• You edit conversation history programmatically and your API account is newer than August 31, 2026. Preserved thinking will reject your requests until the harness goes append-only.
• Sonnet 5 was already good enough. At $2/$10, Sonnet 5 is still half of Opus 5.5’s price with identical cache-read rates, and for routine single-file work the quality gap rarely shows. The upgrade argument is for the hardest 20% of your work, not the routine 80%.
• You need independently verified benchmarks before committing a team. Every score above is vendor-run and one day old. If a standards decision rides on it, wait for the public Terminal-Bench and SWE-bench leaderboards to post independent runs — history says a few points of drift from launch numbers is normal.

And if your real constraint is that any per-token API bill is too unpredictable: an RTX 3090 running Qwen3-Coder locally costs $0/month in tokens (hardware guide on our sister site), and you can trial the workload on a rented 3090 from about $0.07/hr on Vast.ai before buying anything. Open-weight quality sits well below Opus 5.5, but for privacy-bound or budget-capped work it’s the option with no meter running — see the FOSS-tools comparison on aifoss.dev for which harness handles local backends best.

Which model should you actually run now?

Prices as of September 2026, all verified in the comparison above:

Your situation	Run this	Cost	Where
Claude Code on a flat-rate plan	Opus 5.5 (already your default on v2.1.280+)	Included in plan	anthropic.com/claude-code
API-billed agentic coding, quality-first	Opus 5.5	$4 / $20 per MTok	Claude API
API-billed, budget-first	Sonnet 5	$2 / $10 per MTok	Claude API
Cursor user, no BYOK	Opus 5.5 in the picker (Other Models pool)	Plan usage + $4/$20 rates	cursor.com
Cline user today	Pin Opus 5 or Sonnet 5; Opus 5.5 via OpenRouter until the catalog updates	$5/$25 or $2/$10	cline.bot
Still paying for Fable 5.1 coding sessions	Switch to Opus 5.5	60% cheaper per token	Claude API
Privacy-bound / zero-meter	Local Qwen3-Coder on a 24GB GPU	$0/mo tokens; try rented from ~$0.07/hr	Vast.ai

FAQ

Is Claude Opus 5.5 cheaper than Opus 5? Yes, on every line: $4 vs $5 per million input tokens, $20 vs $25 output, $0.20 vs $0.50 cache reads, $5 vs $6.25 cache writes (verified September 23, 2026). Identical token counts cost 20-39% less depending on how cache-heavy the session is; Anthropic additionally claims fewer tokens per completed task.

Does Claude Opus 5.5 replace Opus 5 in Claude Code automatically? Yes, if you’re on Claude Code v2.1.280 or later and ride the default or opus alias — it resolved to Opus 5.5 on September 22, 2026 on every plan and provider except Microsoft Foundry. Pinned model IDs are untouched.

Why can’t I select Opus 5.5 in Cline? Cline’s Anthropic provider only offers models from its built-in catalog, and as of v4.1.20 that catalog predates the launch. Use the OpenRouter provider (live model list, anthropic/claude-opus-5.5) or the Vertex provider (accepts typed model IDs) until a catalog refresh ships — and pin a model regardless, because the unpinned default is Fable 5.1 at $10/$50.

Is Opus 5.5 better than GPT-6 Astra for coding? On Anthropic’s vendor-run table it wins Terminal-Bench 4.0 (66.4% vs 57.9%), FrontierCode (54.4% vs 53.3%), and GDPval-AA (1846 vs 1542), but loses AutomationBench (40.0% vs 41.4%) — and the Terminal-Bench runs used different effort settings per vendor. Treat it as ahead-on-most, not ahead-on-all, until independent leaderboards post.

What happened to the “89.9% SWE-bench Pro” score? It circulates in secondary coverage but does not appear in Anthropic’s published launch table, which uses Terminal-Bench 4.0, FrontierCode v1.1, and CursorBench 4.0 instead. We don’t cite it.

Sources

• Introducing Claude Opus 5.5 — Anthropic announcement
• Claude Opus 5.5 overview — Claude Platform docs (pricing, specs, breaking changes)
• Model configuration — Claude Code docs (default alias resolution, v2.1.280 requirement)
• Claude Opus 5.5 — Cursor docs
• Cline CHANGELOG v4.1.17–v4.1.20 — GitHub (catalog defaults, provider behavior)
• Claude Opus 5.5 is now available in GitHub Copilot — GitHub Changelog
• Claude Opus 5.5 benchmarks explained — Vellum (effort levels, standard error)
• Claude Opus 5.5 launch pricing and benchmarks — Digital Applied

Last verified September 23, 2026. Model pricing and tool integrations change frequently; check the official pages above before committing a team or a budget.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe