Claude Opus 5.5: Pricing, Benchmarks, and Why You Can't Install It Locally
Claude Opus 5.5, announced by Anthropic on September 22, 2026, is the company's new flagship model — and no, you cannot download it, self-host it, or run it on your own hardware, not with a huge GPU cluster, not with anything. Every Claude model Anthropic has ever released, including this one, is closed-weight: it exists only as a hosted service you call over the internet. What you can do is use it today, right now, for a fraction of what the previous flagship cost — Anthropic's own numbers put Opus 5.5 at roughly 40% cheaper to run than Opus 5, while beating it on coding, computer use, and knowledge-work benchmarks. Yes, all this happened in a day from Qwen 2.1 image generation model release.
What Claude Opus 5.5 actually is
A regular at Jake's shop asked him, straight-faced, whether he could "just install the new Claude on the shop laptop" instead of paying for a subscription anywhere. Jake had just watched Ethan explain the exact same misunderstanding about a different AI model days earlier, so this time he already knew the shape of the answer before he even opened a browser tab.
"With some AI companies, 'open' is a real option — they publish the actual model file and you can go get it," Ethan told him. "Anthropic has never done that, for any Claude model, ever. It's not that this one's too big or too new. It's a closed product by design, the way a streaming service doesn't hand you the movie file."
Claude Opus 5.5 is the first release in what Anthropic calls the Claude 5.5 family, and it succeeds Claude Opus 5 as the company's flagship Opus-tier model — the tier built for the heaviest coding, reasoning, and long-running agentic work, as opposed to Sonnet (speed and cost balance) or Haiku (fastest, lightest) tiers. Anthropic's own framing is direct: it performs at roughly the level of Claude Fable 5.1 — the company's most capable model overall — on most everyday work, while costing meaningfully less to run than the model it replaces. Claude Sonnet 5.5 and Claude Haiku 5.5 are confirmed to follow "in the coming weeks," carrying similar improvements down through the rest of the lineup.
One detail matters more than it might seem: Anthropic states this is its first release since the company publicly called for "pacing the frontier" — deliberately not pushing capability faster than its safety practices can keep up with. That framing shows up concretely later in this post, in the safeguards Opus 5.5 launches with.
Your Opus 5.5 request might not actually be answered by Opus 5.5
This is the part almost nobody covering the launch is leading with, and it should be closer to the headline. Anthropic's own safety documentation states that because Opus 5.5 is highly capable at cybersecurity, most cybersecurity-related tasks are automatically re-routed to run on Opus 4.8 instead — transparently, per Anthropic, but nothing in a standard API response tells you which model actually handled a given request. Anthropic's own benchmark footnotes confirm the mechanism runs both ways: "when they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5."
For anyone building a product that bills customers based on which model handled their request, or that logs "answered by Opus 5.5" for compliance or audit purposes, this matters. The model that actually generated a given cybersecurity-adjacent response may not be the one you called by name.
What's already breaking, less than a day in
Two real, sourced early reports worth knowing before you rely on this model in production, both unconfirmed by Anthropic as of this writing:
On Hacker News, within hours of launch, one developer reported that Opus 5.5's max effort setting exhausted its entire 128,000-token output budget while still reasoning about a simple test task (generating an SVG of a pelican riding a bicycle) — producing no actual answer at all. Their conclusion: max effort "may be virtually useless if it's that easy to have it overthink to the point that it doesn't get to a response." One report isn't a pattern yet, but it's a concrete, specific failure mode worth testing before you set any production workflow to max effort by default.
Separately, teams maintaining multi-model routing code have flagged a real integration gap: several open tracking issues describe existing code that pattern-matches on model ID strings failing to recognize claude-opus-5-5 as distinct from claude-opus-5, silently falling through to whatever the "opus" family alias resolves to. If you maintain code that pins a specific Claude model by name rather than by exact version string, this is worth auditing directly rather than assuming your existing regex handles a dotted point release correctly.
Can you install Claude Opus 5.5 locally?
No. Not with a consumer GPU, not with a rented data center, not at all. This is worth stating without hedging because it's a genuinely different situation from an open-weight model that's merely too large for home hardware — there, the weights exist and the barrier is purely physical. Here, there are no weights to acquire in the first place. Anthropic's own documentation on Opus 5.5's system architecture is explicit that it "is not available with 'thinking' mode switched off" and describes the model purely in terms of API access — there is no download, no checkpoint file, no self-hosting path documented anywhere in Anthropic's materials, because none exists.
✅ What "closed-weight" actually gets you here
The trade-off for not being able to self-host is that Anthropic can push cost and safety improvements to everyone using the model simultaneously, the moment they ship — which is exactly what happened with this release. Opus 5.5 arrived with a real price cut already built in, not a separate download you'd need to go find.
What Claude Opus 5.5 actually costs
Pricing is per million tokens, and every line item dropped compared to Opus 5:
| Per 1M tokens | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Cache reads | $0.20 | $0.50 |
| Input tokens | $4.00 | $5.00 |
| Output tokens | $20.00 | $25.00 |
| Cache writes | $5.00 | $6.25 |
Cache reads are the line worth paying attention to: Anthropic notes they make up the majority of cost in long agentic and coding sessions, and that line dropped 60%, not just the standard 20% cut applied to input and output tokens. There's also a Fast mode, available in Claude Code and on the Claude Platform, running up to 2.5× faster at $8 per million input tokens and $40 per million output — a real speed-for-price trade, not a free upgrade.
Anthropic's own framing is that the real-world savings run deeper than the sticker price suggests, because Opus 5.5 also tends to finish the same task using fewer tokens and fewer tool calls than Opus 5 did — the company states this compounds with the price cuts to roughly a 40% total cost reduction on typical workloads, not just a 20% one.
How it compares to Fable 5.1, GPT-6 Astra, and GPT-5.6 Sol
Straight from Anthropic's own published benchmark table:
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench (business workflows) | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity's Last Exam (with tools) | 67.7% | 65.6% | 63.6% | 57.2% | — |
Two honest caveats belong right next to these numbers, both stated by Anthropic itself, not added elsewhere. First, on AutomationBench, GPT-6 Astra edges out Opus 5.5 (41.4% vs 40.0%) — this isn't a clean sweep across every benchmark, and Anthropic doesn't pretend otherwise. Second, and more interesting: Anthropic states plainly that Opus 5.5's scores here were measured with its production safety safeguards active, and when those safeguards intervened during testing, the cybersecurity, biology, and frontier-research tasks in question were completed by older models (Opus 4.8 and Opus 5) instead — which Anthropic says likely reduced Opus 5.5's own measured performance on those specific categories. In other words, the real ceiling on some of these numbers may be a bit higher than what's shown, held down by the model's own safety behavior during the test itself.
Anthropic's own real-world example is a useful sanity check on the "narrower gap than benchmarks suggest" claim: asked to translate HAProxy (widely used load-balancing software) from C into Rust, both Opus 5.5 and Fable 5.1 passed nearly all of HAProxy's own regression tests — Opus 5.5 finished in 9.5 hours against Fable 5.1's 12, at 51% lower cost, but neither model clearly "won" on output quality alone.
The safety restrictions that come with it
What changed between versions
- Before: Opus 5 shipped without the elevated safeguard tier Anthropic reserves for its most capable models — its cybersecurity and biology capabilities didn't trigger the same restrictions.
- Now: Opus 5.5 is the first Opus model to launch with safeguards similar to Fable 5.1's on cybersecurity, biology, and distillation resistance, because Anthropic states its cyber and biology capabilities are now comparable to Claude Mythos 5.1.
- What that means for you: most cybersecurity tasks are automatically re-routed to the older Opus 4.8 model behind the scenes; deep biology research work is gated behind Anthropic's Life Sciences Verification Program for vetted organizations rather than open to everyone by default.
A few other specifics worth knowing rather than discovering mid-project: Opus 5.5 launches with preserved thinking, an anti-distillation safeguard that blocks API users from editing Claude's prior reasoning context to extract it — this applies to any API account created on or after August 31, 2026, and is worth checking against your own account creation date if you're integrating programmatically. It's also, per Anthropic's own documentation, no longer available with "thinking" switched off at all — if you had a workflow built around disabling extended reasoning for speed, that toggle is gone on this model. It does carry EU AI Act watermarking (as Fable 5.1 does) and remains available with zero data retention for accounts that need it.
Where you can actually use Claude Opus 5.5
Live everywhere Anthropic ships models, as of launch day:
- Claude apps — included for Pro, Max, Team, and seat-based Enterprise plans, with raised five-hour usage limits shipping alongside this release.
- Claude Code — selectable directly, including Fast mode for faster (pricier) responses.
- Claude Platform (the API) — model ID
claude-opus-5-5. - Amazon Bedrock — confirmed live the same day, per AWS's own blog.
- Google Cloud (Vertex AI) and Microsoft Foundry (Azure) — both confirmed at launch as well.
Since this site covers AWS regularly, the Bedrock path deserves the specifics. AWS's own documentation confirms Opus 5.5 is billed as a third-party model through AWS Marketplace — charges appear on your AWS bill under the model provider, not a generic Bedrock line item, which is worth knowing if you're reconciling a cost report later. To try it:
- Request model access in the Bedrock console, if you haven't already been granted it for Anthropic models.
- Open Test → Playground and select Claude Opus 5.5 from the model list to try a prompt directly in the console.
- For programmatic access, call it through the Anthropic Messages API shape against
bedrock-runtimeorbedrock-mantle, or stick with the Invoke and Converse APIs via the AWS CLI or SDK if that's what your existing code already uses. - Pick a geo-routing endpoint if data residency matters:
us.anthropic.claude-opus-5-5,eu.,au., andjp.variants each keep requests within that region, or useglobal.for no residency constraint at all.
♂︗ Jake's Reality Check
"So if she can't install it, is there actually anything cheaper than just paying Anthropic directly?"
Not really cheaper, but potentially more convenient depending on what you already use. If your business already runs on AWS, Bedrock bills Opus 5.5 through your existing AWS account rather than a separate Anthropic invoice — same underlying price, one bill instead of two. That's the actual advantage, not a discount.
Frequently asked questions
Is Claude Opus 5.5 free to use?
It's included in Claude Pro, Max, Team, and Enterprise subscriptions within their usage limits, which were raised alongside this launch. Via the API, it's metered per token; there is no unlimited free tier for API access.
Can Claude Opus 5.5 be downloaded or run locally?
No. Anthropic has never released weights for any Claude model. It's available only as a hosted API service, everywhere it's offered.
How much does Claude Opus 5.5 cost via the API?
$4 per million input tokens, $20 per million output tokens, $0.20 per million cached-read tokens, and $5 per million cache-write tokens — all lower than Opus 5's equivalent rates.
Is Claude Opus 5.5 available on Amazon Bedrock?
Yes, confirmed live on launch day per AWS's own blog. It's billed through AWS Marketplace under the model provider rather than a generic Bedrock charge.
How does Opus 5.5 compare to Claude Fable 5.1?
Anthropic states it performs at roughly Fable 5.1's level on most everyday work, while beating it on several benchmarks (Terminal-Bench 4.0, GDPval-AA, AutomationBench) and at meaningfully lower cost per task.
How does Opus 5.5 compare to GPT-6 Astra and GPT-5.6 Sol?
It beats both on most benchmarks Anthropic reports, though GPT-6 Astra edges it out on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science. It's not a universal win across every measure.
What's different between Opus 5.5 and Opus 5?
Lower pricing across every token type, faster output (30%+), fewer tokens and tool calls needed per task, clearer writing style, and a higher safety-safeguard tier matching Fable 5.1 on cybersecurity and biology.
Can you turn off "thinking" mode in Opus 5.5?
No. Per Anthropic's own documentation, Opus 5.5 is not available with thinking mode switched off, unlike some earlier models.
What safety restrictions does Opus 5.5 have?
Cybersecurity-safeguard tasks are largely re-routed to Opus 4.8 automatically; advanced biology work is gated behind Anthropic's Life Sciences Verification Program; and it launches with preserved thinking, an anti-distillation protection, for API accounts created on or after August 31, 2026.
When was Claude Opus 5.5 released?
September 22, 2026, across the Claude apps, Claude Code, the Claude Platform API, Amazon Bedrock, Google Cloud, and Microsoft Azure simultaneously.
What is Opus 5.5's context window?
Anthropic's Opus-tier models in this generation run a 1-million-token context window with a 128,000-token maximum output, consistent with Fable 5.1.
Is Opus 5.5 available on Google Cloud and Microsoft Azure?
Yes, confirmed at launch alongside AWS. Anthropic explicitly states Opus 5.5 is "now available on all platforms."
Does Opus 5.5 support zero data retention?
Yes, available with zero data retention by default on Amazon Bedrock, and offered the same way previous Opus models were on other platforms.
What is Claude Opus 5.5's model ID for the API?
claude-opus-5-5 on the Claude Platform directly; on Amazon Bedrock, geo-specific variants like us.anthropic.claude-opus-5-5 or global.anthropic.claude-opus-5-5 are also available depending on data-residency needs.
Will there be a Claude Sonnet 5.5 or Haiku 5.5?
Anthropic has confirmed both are coming "in the coming weeks," carrying many of the same performance, efficiency, and safety improvements introduced with Opus 5.5.
Is Opus 5.5 good for coding?
This is Anthropic's strongest stated area for the model — leading on Terminal-Bench 4.0, FrontierCode, and CursorBench, with real-world examples like a 200,000-line codebase audit finishing in under 3 hours versus over 20 hours for Opus 5.
Revision note. Written the same day Claude Opus 5.5 launched, September 22, 2026, directly from Anthropic's own release page and AWS's own Bedrock announcement. Pricing, benchmark scores, and safeguard details for a same-day release are the parts most likely to be refined in the days after launch folks so— check Anthropic's pricing page directly before treating any number here as permanent. If you came here after hearing "the new Claude" was free and open, you're not the only one — that mix-up seems to be going around this week.