Claude Opus 5.5: What the Price Cut Really Buys You

Claude Opus 5.5 cut list prices by 20% and matches Fable 5.1 on most benchmarks. What changed, what broke, and who should actually pay for it.

Disclosure: this article is built from Anthropic’s published announcement, pricing pages, system card coverage and third-party benchmark reporting, not from a long-running production deployment. Figures are attributed in the text.

Claude Opus 5.5 arrived on September 22, 2026, and the headline is simple: it costs less than Opus 5 and, by Anthropic’s own account, performs at roughly the level of Claude Fable 5.1 on most work. That combination is rare. Price cuts usually come with a quiet downgrade, and capability jumps usually come with a bigger bill.

The honest answer up front: Claude Opus 5.5 is the right model for long, judgment-heavy agent work, and the wrong default for almost everything else. Six days after it launched, Sonnet 5.5 landed at half the price and within a few points on most coding tests. So the real question is not whether Opus 5.5 is good. It is whether your workload is one of the minority where the gap is worth paying for.

Table of Contents

Claude Opus 5.5 workspace with notebook and laptop
The price per token fell 20%, but the bigger saving comes from Claude Opus 5.5 using fewer tokens to finish the same job.

What Claude Opus 5.5 Actually Is

Claude Opus 5.5 is the first model in Anthropic’s Claude 5.5 family. It replaced Opus 5, which had only been out since July 24, 2026, so the turnaround was about two months. Sonnet 5.5 followed on September 28 and Haiku 5.5 on October 7, which means the whole family refreshed in roughly a fortnight.

The model ID is claude-opus-5-5, with no date suffix. On Amazon Bedrock it shows up as anthropic.claude-opus-5-5. It is also live on Google Cloud, Microsoft Foundry and Claude Platform on AWS, so if your company is locked into one cloud for procurement reasons, you do not have to route around it.

Anthropic does not publish parameter counts, and it did not here either. What it did publish is a context window of one million tokens billed at standard rates across the whole window, a maximum output of 128,000 tokens, and a knowledge cutoff of June 2026. The model outputs text only, though it reads images.

The positioning matters. In the current lineup, Fable sits at the top, then Opus, Sonnet and Haiku. Claude Opus 5.5 is smaller and cheaper than Fable 5.1, yet the published benchmark table puts it ahead of Fable 5.1 on most rows. That is the part that made people sit up, because it means the second tier now beats last month’s top tier for most practical jobs.

There is a commitment worth knowing about too. Anthropic says it will keep Claude Opus 5.5 available until at least September 22, 2027. If you have ever rebuilt a prompt pipeline because a model was retired on short notice, a guaranteed year is a real feature.

Claude Opus 5.5 Pricing, Line by Line

The list price is $4 per million input tokens and $20 per million output tokens. Opus 5 was $5 and $25, so that is a straight 20% cut. Cache reads dropped harder, from $0.50 to $0.20 per million, which is a 60% reduction and probably the most important number for anyone running agents.

Item Opus 5 Claude Opus 5.5
Input (per 1M tokens) $5 $4
Output (per 1M tokens) $25 $20
Cache reads (per 1M) $0.50 $0.20
5-minute cache writes (per 1M) $6.25 $5
Context window 1M 1M, no long-context surcharge

Here is where the marketing and the arithmetic diverge. Anthropic says Claude Opus 5.5 is about 40% cheaper to run than Opus 5, not 20%. The extra saving comes from the model finishing tasks in fewer tokens. That is plausible and some early reports back it up, but it depends entirely on your workload. A summarisation job with fixed-length output will see close to the 20% list cut. A long agent loop that used to wander might see much more.

Why cache reads matter so much: an agent session re-sends its system prompt, tool definitions and conversation history on every turn. On a 100-turn session, most of the input is cached context. Cutting that rate by 60% changes the economics of long-running agents more than the headline input price does.

Opus 5.5 pricing still sits well above the competition at the mid tier. Sonnet 5.5 and GPT-6.1 Sol both list at $2 and $10. So you are paying exactly double for Opus over those two. Whether that is a bargain or a waste depends on the next section.

How Claude Opus 5.5 Benchmarks Against Rivals

Anthropic published a nine-row comparison against Fable 5.1, Opus 5, GPT-6 Astra and GPT-5.6 Sol. Several outlets transcribed it in full. The pattern is consistent: Claude Opus 5.5 leads or ties on agentic coding and long-horizon tasks, and trails in a few knowledge-heavy rows.

One claim stood out in coverage of the launch. Anthropic says Claude Opus 5.5 beats GPT-6 Astra’s max-effort GDPval-AA v2.1 score at about a fifth of the cost per task. GDPval measures performance on realistic professional work, so that is a meaningful number, even if it is vendor-reported.

Independent data tells a more mixed story. On factual recall, Artificial Analysis’s Omniscience test puts Opus 5.5 at 66%, comfortably ahead of Sonnet 5.5 at 54%. That gap is real and it shows up in research and analysis work, where knowing things matters as much as reasoning about them.

On cost per task, though, the competition is fierce. DataCamp’s write-up of GPT-6.1 Sol quotes OpenAI’s figures on Terminal-Bench Science at max effort: $5.47 per task for Sol, $23.21 for Opus 5.5 and $23.80 for Astra. OpenAI also reports Sol 2.2 points above Opus 5.5 on AutomationBench at medium effort, at about a third of the cost. Those are OpenAI’s numbers, so treat them with the same scepticism you would apply to Anthropic’s.

My read: on raw capability, Claude Opus 5.5 is a frontier model. On capability per dollar, it is not the leader. That is fine. Frontier models have never been the efficient choice; they are the choice when the cheaper option fails often enough to cost you more in retries and review time.

The Breaking Changes Nobody Put in the Headline

Several guides counted four breaking changes against Opus 5. The one most likely to bite you is that extended thinking can no longer be fully disabled. Thinking is always on. If you built latency-sensitive features on Opus 5 with thinking switched off, you will need to retest, because you cannot simply flip the old setting back.

Effort levels replace manual thinking budgets as the main dial. Practitioners report that the right effort setting for Opus 5 is not necessarily the right one for Claude Opus 5.5, so carrying your old configuration over unchanged is the most common migration mistake.

The New Stack’s launch coverage flagged another wrinkle in its headline: some agent calls might get routed to an older model. The detail concerns how Anthropic handles certain safety-flagged requests, which can fall back rather than fail. For most users this never surfaces. For anyone running evaluations, it is a reason to log which model actually answered each request.

There are also stricter safeguards. VKTR’s coverage describes Opus 5.5 as shipping with Fable-level cyber and biology protections. If your work involves security research, expect more refusals than you saw on Opus 5 and plan for a review path.

Claude Opus 5.5 vs Sonnet 5.5

This is the comparison that actually matters for most budgets. Sonnet 5.5 launched on September 28 at $2 and $10, half the price of Claude Opus 5.5 on every line. Anthropic’s own framing is that Sonnet comes close to Opus on most benchmarks while Opus remains clearly stronger at complex, open-ended work that requires sustained judgment.

Dimension Sonnet 5.5 Claude Opus 5.5
Input / output per 1M $2 / $10 $4 / $20
Cache reads per 1M $0.10 (from October 7) $0.20
AA-Omniscience (factual recall) 54% 66%
Context 1M 1M
Best fit Daily coding, agents, volume work Long-horizon, judgment-heavy tasks

DataCamp noted that Sonnet 5.5 even edges past Opus on one terminal coding benchmark. Claude Fast priced a 100-turn agent session at about $4 on Sonnet against $8 on Opus at current cache rates. Double the money for a few points of accuracy is a good deal only if those points are the difference between a task that works and one that needs a human to rescue it.

A practical way to decide: run your twenty hardest real tasks through both. Count the ones Sonnet fails that Claude Opus 5.5 completes. If that number is close to zero, Sonnet is your default and Opus is your escalation path.

Where Claude Opus 5.5 Earns Its Price

There are three kinds of work where I would reach for it first.

Long agent runs with many dependent steps

Errors compound. A model that is 95% reliable per step is roughly 60% reliable over ten steps. A few points of per-step reliability turn into a large difference in whole-task success. This is where Claude Opus 5.5 justifies itself: multi-hour refactors, migrations across a large codebase, research tasks that chain dozens of tool calls.

Work where being wrong is expensive

Contract review, financial analysis, anything a person will act on without checking every line. The factual-recall gap over Sonnet matters here, and so does the steadier judgment Anthropic keeps emphasising.

Writing that has to sound like a person

Anthropic promised a clearer, more natural voice with Opus 5.5, and The New Stack noted that Sonnet 5.5 later picked up the same trait. If prose quality is the product, both are improvements over the 5.0 generation, but Opus still handles long documents with fewer drifts in tone.

Where it does not earn its price: classification, extraction, routing, short summaries and simple chat. Haiku 5.5 at $0.10 input does those jobs for a fortieth of the cost.

Migrating to Claude Opus 5.5 Without Surprises

If you are moving from Opus 5, the swap is mostly a model-name change, but a few checks save pain later.

  • Retest effort levels. Start at medium and move up only for tasks that fail. Higher effort spends more output tokens, which are the expensive ones.
  • Remove code that disables thinking. It will no longer do what you expect.
  • Restructure prompts for caching. Put stable content first. The minimum cacheable prompt is 512 tokens, so short system prompts may not cache at all.
  • Log the responding model. If routing falls back for some requests, you want to know.
  • Re-baseline costs after a week. The 40% figure is an average. Yours will differ.

On the subscription side, Claude Code changed its default for Pro and Team Standard users to Opus, according to the changelog cited in Penchan’s guide. Free-plan users still cannot select Claude Opus 5.5. Pro, Max, Team and Enterprise subscribers can pick it from the model menu, and they also received a banked rate-limit reset to save for heavy days.

What the Two-Month Release Cycle Means for You

Opus 5 lasted about eight weeks as the flagship general-release model. That pace is new, and it changes how you should build. A year ago, picking a model was a decision you made once a quarter. Now it is closer to a monthly review, and the teams that cope best are the ones that treat the model name as configuration rather than something baked into code.

In practice that means three habits. Keep a small, versioned evaluation set of real tasks with known good answers, so you can score a new release in an afternoon instead of arguing about anecdotes. Put model selection behind one function or one config value, so a swap is a single change and a rollback is just as quick. And record cost and success per task type, because the right answer for summarisation is rarely the right answer for refactoring.

The fast cadence also argues for caution with fine-grained prompt tuning. Prompts tuned obsessively for one model often underperform on its successor, because the new model follows instructions more literally or needs less hand-holding. Anthropic said the 5.5 generation writes more clearly and needs fewer tokens to finish a task, which is exactly the kind of change that makes old scaffolding counterproductive. Simple, explicit prompts age better than clever ones.

Finally, watch the retirement dates rather than the launch dates. A model you depend on is only as useful as its guaranteed lifetime. The one-year commitment here is generous by industry standards, but it is still finite, so put a reminder in the calendar for next summer and plan the next migration before you are forced into it.

Who Should Use What

If you are Start with Why Link
Running long autonomous coding agents Claude Opus 5.5 Per-step reliability compounds over many steps See Claude models
A team doing daily coding and chat Sonnet 5.5 Half the price, within a few points See Sonnet 5.5
Processing high-volume simple tasks Haiku 5.5 $0.10 input, built for volume See pricing
Optimising strictly for cost per task GPT-6.1 Sol Strong vendor-reported cost per task Read the specs

Related on this site: for more context, see Claude Sonnet 5.5, GPT-6.1 Sol and Claude Fable 5.2 vs Mistral Large 4 vs Kimi K3.1 Pro.

Frequently Asked Questions About Claude Opus 5.5

When was Claude Opus 5.5 released?
Anthropic released it on September 22, 2026, as the first model in the Claude 5.5 family. Sonnet 5.5 followed on September 28 and Haiku 5.5 on October 7.

How much does Claude Opus 5.5 cost?
It lists at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. That is 20% below Opus 5 on input and output, and Anthropic says real costs fall about 40% because the model uses fewer tokens.

Is Claude Opus 5.5 better than Fable 5.1?
Anthropic says it performs at roughly Fable 5.1’s level on most work and beats it on several agentic coding benchmarks. Fable models remain Anthropic’s top tier, so for the hardest problems the Fable line is still the ceiling.

Can I turn off extended thinking?
No. Thinking is always on in Claude Opus 5.5, and effort levels are the control you use instead. Teams that disabled thinking on Opus 5 for speed need to retest latency.

What is the context window?
One million tokens, billed at the standard rate across the whole window, with up to 128,000 output tokens. There is no long-context surcharge, unlike some competitors that double prices past 272,000 tokens.

Should I use Claude Opus 5.5 or Sonnet 5.5?
For most teams, Sonnet 5.5 should be the default and Opus the escalation path. Test your hardest real tasks on both and pay for Opus only where Sonnet actually fails.

Verdict

Use Claude Opus 5.5 for the work where a failure costs more than the token bill: long agent runs, high-stakes analysis and anything that chains many dependent steps. For those jobs, the price cut and the cheaper cache reads make it the best value Opus has ever been.

The honest caveat is that Sonnet 5.5 arrived six days later at half the price and closed most of the gap. Benchmarks from both Anthropic and OpenAI are vendor-reported, and the 40% saving is an average that your workload may not match. Treat every number here as a starting hypothesis.

The concrete next step: pick twenty real tasks from last month, run them through Sonnet 5.5 and Claude Opus 5.5 at medium effort, and record success rate and cost. Route by the result, not by the leaderboard.

Sources and Further Reading

On this site: compare the top tier in our Claude Fable 5.1 guide, cut your bill with LLM API cost optimization, pick a coding tool from AI coding assistants in 2026, and see the wider field in AI models compared.

2 Comments

Comments are closed.