September 22, 2026 ·

anthropicclaudepricingopuseffortprocurementsafeguardsmigration

Claude Opus 5.5 is 20% cheaper on the rate card and 40% cheaper in the press release — the gap is a default setting

TL;DR: Anthropic released Claude Opus 5.5 on 22 September 2026 at $4 input / $20 output per million tokens, down from Opus 5’s $5/$25, with cache reads cut 60% from $0.50 to $0.20. The headline claim is that it “costs 40% less to run than Opus 5.” The rate card explains roughly half of that; the rest comes from the effort parameter’s default dropping from high on Opus 5 to medium on Opus 5.5 — fewer thinking tokens by default, on a model that at equal effort thinks more than its predecessor. Separately, flagged requests do not run on the model you chose: cyber-flagged work falls back to Opus 4.8, biology and frontier-LLM work to Opus 5 — both billed at $5/$25, i.e. 25% above the rate of the model you selected. On the API this surfaces as a refusal you have to handle, not a silent swap. Four breaking changes make this more than a model-ID edit. Specs: 1M context, 128K max output, Jun 2026 knowledge cutoff, claude-opus-5-5.

What shipped

Claude Opus 5.5 went live on 22 September 2026 across the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, as claude-opus-5-5 (anthropic.claude-opus-5-5 on Bedrock). It is a 1M-token-context model with a 128K max output — 300K on the Batch API behind the output-300k-2026-03-24 beta header — taking text and images in, text out, with a June 2026 knowledge cutoff. Retirement is committed at not sooner than 22 September 2027.

The positioning is unusual for a frontier launch: the lead claim is about cost, not capability. Anthropic says the model “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5,” generates output more than 30% faster, and — addressing a long-running complaint — “puts the most important information first, uses less jargon, and follows writing instructions more closely.”

Reported benchmark results:

Benchmark	Opus 5.5	Opus 5	Fable 5.1
Terminal-Bench 4.0	66.4%	52.3%	55.8%
FrontierCode v1.1	54.4%	48.0%	50.3%
OSWorld 2.0	81.8%	74.0%	80.7%
GDPval-AA v2.1 (Elo)	1846	1708	—
Chartography	89.0%	—	88.4%

Taken at face value, that is a model beating a flagship priced at $10/$50 on agentic coding while charging $4/$20. Pre-release evaluation was run externally by Frontier Design and METR.

The 20% that is unconditional

The rate card moved, and this part needs no configuration and no argument:

	Opus 5.5	Opus 5	Change
Input	$4 / MTok	$5 / MTok	−20%
Output	$20 / MTok	$25 / MTok	−20%
Cache read	$0.20 / MTok	$0.50 / MTok	−60%
5m cache write	$5 / MTok	$6.25 / MTok	−20%
1h cache write	$8 / MTok	$10 / MTok	−20%
Batch	$2 / $10	$2.50 / $12.50	−20%
Fast mode	$8 / $40	$10 / $50	−20%

The cache-read line is the interesting one and it is being under-reported. Opus 5.5 prices cache hits at 0.05x the base input rate rather than the standard 0.1x — the same trick Anthropic used when it cut Fable 5.1’s cache reads to 0.025x on 1 September. For a long-running agent that replays a large system prompt and a growing transcript on every turn, cache reads are the bulk of the bill, and this is a 60% cut on the bulk of the bill.

Work an example. A turn that reads 1M cached tokens, adds 100K fresh input and emits 20K output:

Opus 5: $0.50 + $0.50 + $0.50 = $1.50
Opus 5.5: $0.20 + $0.40 + $0.40 = $1.00

That is a 33% saving with no behavioural change at all. Push the shape the other way — little caching, lots of output — and the same arithmetic converges on the flat 20%. So the rate card alone delivers somewhere between 20% and about a third, depending on workload shape, which is exactly the range we mapped when Fable 5.1’s cache pricing moved.

Which leaves a gap to 40%.

The other half is a default, not a discount

Anthropic’s own phrasing closes the gap, and it is precise rather than evasive: “Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads.”

The default settings changed. From the model comparison table in Anthropic’s own docs:

Model	Price / MTok	Default effort
Claude Fable 5.1	$10 / $50	`high`
Claude Opus 5.5	$4 / $20	`medium`
Claude Opus 5	$5 / $25	`high`
Claude Sonnet 5	$2 / $10	`high`

Opus 5 defaulted to high. Opus 5.5 defaults to medium. Thinking tokens are billed as output tokens, so a lower default effort is a smaller bill for the same prompt — not because the tokens got cheaper, but because there are fewer of them. The migration notes say so plainly: “A request that omits effort runs at medium; on Claude Opus 5 it ran at high. Set effort explicitly and re-run your sweep.”

This is not a trick, and the model is genuinely better per token — a 14-point gain on Terminal-Bench 4.0 at a lower effort default is a real result. But it means the 40% and the 20% describe two different products. A team that inherits the default gets a cheaper, faster, slightly less deliberative model. A team that pins effort: high to hold its existing quality bar gets the rate-card cut and roughly none of the rest.

There is a sting in the tail for that second team. From the same behaviour notes: “At the same effort setting the model tends to think more per turn than Claude Opus 5, most of all at xhigh and max.” So pinning effort back to high does not restore Opus 5’s token consumption — it overshoots it. It is entirely possible to migrate to a model with a 20% lower rate card and watch the invoice go up.

This is the third launch this month built on the same structure. Grok 4.7 held its $2/$6 rate card and moved the benchmark into a higher effort column; Claude Code’s Projects rollout shipped an opinionated effort default of its own. The per-token price is becoming the least informative number on a launch page, and the effort setting the most.

Three models wearing one model ID

The second story here is what happens when a request touches a restricted domain. Opus 5.5 runs a biology safety classifier in addition to the cybersecurity one, plus a new reasoning_extraction category for requests that push the model to reproduce its internal reasoning as response text.

Anthropic’s launch page says “most cybersecurity tasks will be re-routed to Opus 4.8,” and reporting on the launch adds that biology and frontier-LLM-development requests route to Opus 5. Some coverage has framed this as agent calls being “secretly” downgraded. On the Claude API, that framing is wrong, and the correction matters because it changes what a team has to build.

What the API actually does on a flagged request is refuse, visibly. The response is HTTP 200 with stop_reason: "refusal" and a stop_details.category naming the policy area — one of cyber, bio, frontier_llm, reasoning_extraction or general_harms. No substitution happens unless fallback is explicitly enabled by setting fallbacks: "default" and sending the server-side-fallback-2026-07-01 beta header. And when it is enabled, Anthropic documents the handoff as visible: “the response names the model that served it, and the fallback content block marks the handoff,” with a per-attempt billing record in usage.iterations.

A corroborating detail: Anthropic’s help-centre article on real-time cyber safeguards carries an explicit carve-out — “This article applies only to Opus and Sonnet class models, but doesn’t apply to Claude Opus 5.5.” Opus 5.5 is on the refusal-and-fallback mechanism, not the older block-in-place one.

So the honest reading is split by surface. In the Claude apps and consumer surfaces, the re-route happens on the user’s behalf. On the API, the default is a refusal that an integration must handle, and the substitution is opt-in and labelled. What is genuinely easy to miss is the price:

Refusal category	Fallback model	Fallback price
`cyber`	Claude Opus 4.8	$5 / $25
`bio`	Claude Opus 5	$5 / $25
`frontier_llm`	Claude Opus 5	$5 / $25

The fallback costs more per token than the model that was selected. A refusal arriving before any output is not billed, but every attempt that produced output is billed at the rates of the model that ran it. A security team whose legitimate defensive work trips the cyber classifier regularly is not getting a cheaper model at a lower rate — it is getting an older model at a 25% higher rate, twice the round trips, and a capability ceiling it did not choose.

The escape hatch is a programme, not a purchase. Anthropic says it will expand its Cyber Verification Program to cover Opus 5.5 “in the coming weeks,” with three tiers of increasingly permissive trusted access, up to and including Claude Mythos models. Biology work runs through the Life Sciences Verification Program and its retention terms. This is now the industry’s settled shape: the rate card buys the restricted model, and full capability is a clearance rather than a line item.

The migration is not a model-ID edit

Four changes return errors on code that runs today against Opus 5:

Thinking cannot be disabled. thinking: {"type": "disabled"} and manual budget_tokens both return a 400. Effort is the only lever.
Forced tool use is gone. tool_choice of type any or tool returns a 400. Schema-constrained calls move to strict tool use or structured outputs.
Thinking blocks are bound to their producing model. Opus 5.5 reads blocks from Opus 5 and earlier Opus, Sonnet and Haiku models, but not from Claude Fable or Claude Mythos. Blocks the target cannot read are dropped, not billed, and reported in input_transformations under a beta header.
computer_20251124 is rejected on the Claude API and Google Cloud; the computer_toolset_20260801 toolset is required. On Bedrock the old tool still works — a platform split worth checking before assuming parity.

And one that fails nothing, which is why it will bite hardest: the short notes the model writes between tool calls now come back as thinking blocks rather than text blocks, with empty text at the default display: "omitted". An agent UI that streams those as progress updates simply goes quiet between tool calls. No error, no alert, just a product that looks hung. This is the same class of problem as the usage fields that stopped describing the bill after on-demand compaction shipped, and the gateway field-drop that quietly changed what Claude Code charged for: a response-shape change that passes every test and breaks the thing users actually see.

What to do in the next two weeks

Collect the unconditional saving first. Cache reads dropped 60%. If a workload is cache-heavy, that money arrives on a model-ID change with no tuning, and it is the largest single line in most agent bills.

Pin effort explicitly and re-run the sweep. Do not inherit medium by accident and do not carry high across by habit. Both are now decisions with a price attached, and at xhigh and max the new model consumes more than the old one at the same label.

Measure cost per completed task, not cost per million tokens. Every number in this article is a rate. The only figure that settles a migration is the total spend to finish one representative internal workload, run twice. This is the same discipline the Grok 4.7 launch demanded, and it is becoming the only defence against rate cards that no longer describe invoices.

Handle stop_reason: "refusal" before you ship. If an integration branches only on end_turn and max_tokens, a refusal is an HTTP 200 with empty content — a silent null result, not an exception. Branch on stop_reason, not on content, and note that stop_details.category can legitimately be null.

Budget flagged work at $5/$25. Any team whose prompts routinely touch security, biology or model-training topics should price the fallback, not the headline, and start the Cyber Verification Program application now rather than after the first invoice.

Check the progress-update path. If your interface streams inter-tool text to users, set thinking.display before migrating, or ship a product that goes mute mid-task.

The pattern

Three of this month’s four significant launches have decoupled the advertised price from the delivered price, each by a different mechanism. Grok 4.7 held the rate card and raised the effort column. Xiaomi’s MiMo-V2.6-Pro undercut the field by an order of magnitude and moved the argument to licensing. Opus 5.5 cuts the rate card honestly and then books a further saving against a default it changed in the same release.

None of these is deceptive. Anthropic wrote “at default settings” in the sentence, documented the effort change in the migration notes, and published the fallback categories and their billing rules in full. The problem is structural rather than ethical: a buyer comparing two models now has to reconcile a per-token rate, an effort default, a token-efficiency delta at matched effort, a cache-read multiplier, a classifier fallback with its own rate card, and a tokenizer that changed under both models two versions ago. No spreadsheet built from price lists survives that. Only a replayed workload does.

The genuinely good news is narrower than the headline and more durable. Opus 5.5 beats a $10/$50 flagship on agentic coding at $4/$20, cuts the single largest line in an agent bill by 60%, and commits to a year of availability. That is worth migrating for. Just measure it yourself, because the number on the announcement page is measuring a configuration, not a model.

For where this leaves the wider field, see the Claude review, the Claude Code review, the Claude vs ChatGPT comparison and the best AI coding tools roundup. Subscription users should read this alongside the weekly limit changes from earlier this month, which price the same capability on a different meter entirely.

Update, 23 September 2026 — the model that launched ninety minutes later costs half as much. This article’s closing point was that Opus 5.5 beats a $10/$50 flagship on agentic coding at $4/$20. On the same day, roughly ninety minutes after Anthropic, OpenAI shipped GPT-6 Sol at $2/$10 — half Opus 5.5’s rate card on both halves, with identical $0.20 cached input.

The matched-effort comparison still favours Anthropic, and it favours it in a way worth pinning down. On Artificial Analysis’ Intelligence Index v4.3.2 at max effort, Opus 5.5 scores 57.6 against Sol’s 47.5, and leads on GDPval-AA v2.1 Elo (1,846 to 1,487), terminal tasks (59.6% to 43.9%) and SciCode (66.9% to 57.6%). The decisive row is the one neither vendor published: Opus 5.5 at medium scores 51.2 at ~$1.34 per task, against Sol at max at 47.5 for ~$1.06 — 3.7 points more for about 26% more money, on roughly a fifth fewer output tokens. Opus at max reaches 57.6 but costs ~$5.98 per task, a 4.5x spend for 6.4 points.

Which sharpens the advice above rather than changing it: pin effort deliberately, then measure cost per completed task. The cheaper model at its most expensive setting is currently losing to the pricier model at its cheapest one.

Frequently asked questions

Is Claude Opus 5.5 actually 40% cheaper than Opus 5?

It depends on whether the effort setting is inherited or pinned. Anthropic's wording is precise and worth reading twice: 'at default settings it will cost 40% less than Opus 5 on typical workloads.' The rate card contributes a 20% cut to base input and output ($5/$25 to $4/$20) and a 60% cut to cache reads ($0.50 to $0.20 per million). On a cache-heavy agent loop those two together land around 30-33%; on an output-dominated workload they land near 20%. The remaining distance to 40% comes from the model emitting fewer tokens, and the largest single reason it emits fewer tokens is that the effort parameter's default moved from high on Opus 5 to medium on Opus 5.5. A team that sets effort explicitly to high to preserve its existing quality profile keeps the rate-card saving and gives back most of the rest. Both numbers are honest; they describe different configurations.

What happens if my request gets flagged as a cybersecurity task?

On the Claude API it does not silently swap models. A declined request returns HTTP 200 with stop_reason set to 'refusal' and a stop_details object whose category field names the policy area — cyber, bio, frontier_llm, reasoning_extraction or general_harms. Nothing is routed anywhere unless fallback is switched on. Server-side fallback is opt-in: set the fallbacks parameter to 'default' and send the server-side-fallback-2026-07-01 beta header, and the API retries on the model Anthropic recommends for that category. Anthropic's documentation is explicit that this is visible — 'the response names the model that served it, and the fallback content block marks the handoff.' The re-routing described on the launch page as 'most cybersecurity tasks will be re-routed to Opus 4.8' is the behaviour on the Claude apps and consumer surfaces, where the decision is made for the user. API integrations that never set the fallbacks parameter get a refusal to handle, not a quiet downgrade.

Does the fallback model cost less, since it is an older model?

No — it costs more per token, which is the counterintuitive part. Opus 5.5 is $4 input and $20 output per million tokens. Opus 4.8, the recommended fallback for the cyber category, is $5 and $25. Opus 5, the fallback reported for biology and frontier-LLM-development flags, is also $5 and $25. So a request that falls back is billed 25% above the rate of the model that was actually selected. Anthropic's billing rules are clear about the mechanics: an attempt that declined before producing any output is not billed at all, while every attempt that produced output is billed separately at the rates of the model that ran it, with the per-attempt record in the usage.iterations array. A security team running high volumes of flagged-but-legitimate work should model its spend at $5/$25, not $4/$20, until it is accepted into the Cyber Verification Program.

What will break if I just change the model ID?

Four things return errors and one changes the response shape without failing. Requests that set thinking to disabled, or that set a manual budget_tokens, now return a 400 — thinking is always on and effort is the only control. Forced tool use is gone: tool_choice of type 'any' or 'tool' returns a 400, so schema-constrained calls need strict tool use or structured outputs instead. Thinking blocks are bound to the model that produced them, and Opus 5.5 cannot read blocks from Claude Fable or Claude Mythos models. On the Claude API and Google Cloud the older computer_20251124 computer use tool is rejected and the computer_toolset_20260801 toolset is required, though on Amazon Bedrock the old tool still works. The silent one: the short notes the model writes between tool calls now arrive as thinking blocks rather than text blocks, so an application that streams them as progress updates goes quiet between tool calls with no error until it sets a thinking display value.

Should I move off Opus 5 or Fable 5.1 this week?

Opus 5 is not going anywhere — it is still listed at $5/$25 and Opus 5.5 carries a retirement commitment of not sooner than 22 September 2027, so there is no forced-migration clock. The case for moving is strongest for cache-heavy agent loops, where the cache-read cut from $0.50 to $0.20 is unconditional and needs no configuration change to collect. The case for moving off Fable 5.1 is different and larger: Fable 5.1 is $10/$50, and Opus 5.5 reportedly beats it on Terminal-Bench 4.0 (66.4% against 55.8%) and FrontierCode v1.1 (54.4% against 50.3%) at 40% of the price. That is the comparison Anthropic wants read and it is also the one with the most money in it. The thing not to do is carry an effort setting across the migration unexamined — at any given effort level Opus 5.5 tends to think more per turn than Opus 5, most of all at xhigh and max, so a team that pins effort high can end up paying more per task on a model with a lower rate card.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.