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 Opus-tier 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. Here's the counterintuitive part almost nobody leads with: the model that beats Anthropic's bigger, more expensive flagship on most benchmarks also costs roughly 40% less to run than the model it replaces — $4 per million input tokens and $20 per million output, down from $5 and $25 — and it arrived with its safety restrictions turned up, not down, as the first release under Anthropic's deliberate "pace the frontier" policy.
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 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."
What Claude Opus 5.5 Actually Is
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 system is worth thirty seconds because it's how every Claude decision gets made: Opus is the tier built for the heaviest coding, reasoning, and long-running agentic work — "agentic" meaning AI that plans multi-step tasks and uses tools along the way, rather than answering once and stopping. Sonnet sits in the middle, trading some capability for speed and lower cost. Haiku is the fastest, lightest tier. Above Opus sit Anthropic's even bigger models, Fable and Mythos — and Anthropic claims Opus 5.5 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 Opus 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 — so if you're choosing a Claude plan today, know that the mid-tier is about to shift under you.
One piece of context matters more than it might seem: this is Anthropic's first release since CEO Dario Amodei publicly embraced what the company calls pacing the frontier — deliberately not pushing capability faster than safety practices can keep up. Amodei's words, from a post earlier in the month: "I have become convinced that fully addressing the risks requires even more prudence... not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up." That framing isn't press fluff; it shows up concretely in the safeguards Opus 5.5 launches with, which we'll get to — and it explains why this release leads with cost efficiency and safety tiering rather than a raw capability jump.
Your Opus 5.5 Request Might Not 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. 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, Anthropic claims, 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 the safeguards intervened during testing, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5.
⚠️ What this actually breaks
If you build 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, the model that actually generated a given cybersecurity-adjacent response may not be the one you called by name. There is no flag in the standard response telling you a re-route happened. If your work touches security tooling, penetration testing, exploit analysis, or anything adjacent, assume some percentage of your Opus 5.5 calls are secretly Opus 4.8 calls — priced the same, benchmarked separately, and not labeled.
Jake's read on it was the honest one: "So I order the specialist, and the company quietly sends the junior guy for some jobs — same bill." That's a fair compression. The re-routing exists for defensible safety reasons — the safeguard tier exists because the capability is real — but "defensible" and "labeled in your logs" are different things, and only one of them is true today.
The Effort Dial, and Why Max Can Backfire
Opus 5.5 supports a ladder of effort settings running from low up to max, controlling how much internal reasoning the model does before it answers. Low effort responds fast with minimal thinking; max effort lets the model reason for extended stretches on hard problems. The change worth knowing: on Opus 5, thinking could be switched off entirely at lower effort levels via the API — on Opus 5.5, thinking is always on at every effort level, including on the API. There is no "off" switch. If you built a workflow around disabling extended reasoning for speed, that toggle is gone on this model, and every response carries some reasoning overhead whether the task needs it or not.
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: on a model where thinking cannot be disabled, the highest effort setting can spend the whole budget on thinking. Test before you set any production workflow to max by default.
Separately, teams maintaining multi-model routing code have flagged a real integration gap: existing code that pattern-matches on model ID strings can fail 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, audit it directly rather than assuming your existing pattern handles a dotted point release.
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.
A quick vocabulary check, because two words get swapped constantly: a model's weights are the learned numbers inside it — the actual file, the thing you'd need to run it yourself. An open-weight model publishes that file for anyone to download and run. A closed-weight model never does. Anthropic has never released weights for any Claude model, and Opus 5.5 is no exception: there is no download, no checkpoint file, no self-hosting path documented anywhere in Anthropic's materials, because none exists. The model is a service you call, the way Netflix is a service you stream.
"So the people online selling 'Claude offline installers'..." Jake started asking, and Ethan finished the sentence: "Are selling you something else with Claude's name on it. Every single one. If it runs without an internet connection, it is not Claude."
✅ What closed-weight actually gets you here
The trade for not being able to self-host is that Anthropic can push cost and safety improvements to everyone 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. If your priority is privacy-by-hardware or zero API spend, that trade is wrong for you — but it's the only trade on offer.
What You Can Run Locally Instead
Since the local alternative to Claude isn't "Claude, smaller" — it's a different model entirely — it's worth being precise about what the honest option is. The tool most people use is Ollama, a local model runner, and the reason it comes up in Claude conversations specifically is that Ollama supports a subset of the Anthropic Messages API — enough for agentic coding tools like Claude Code to be pointed at a local, open-weight model instead of Anthropic's servers. Messages, streaming, and function calling work; tool-choice controls and some advanced features don't. That's the honest boundary
- Install a local model runner — Ollama is the one most people use, and it runs on ordinary hardware.
- Download an open-weight model that fits your RAM — the models people actually run this way (GLM, DeepSeek, Gemma, and similar) are open-weight, which is precisely what Claude is not.
- Point a tool that speaks the Anthropic Messages API at it — with the endpoint swapped to your local runner, the tool talks to your model instead of Anthropic's.
The honest trade, stated plainly: you give up the benchmark performance in the table below, the 1M-token context window, and the always-on safety training — and you gain privacy (nothing leaves your hardware) and cost (zero per token, you already own the electricity). For most people asking "should I run Claude locally," the answer is "no, and here's what you're actually asking" — but for the specific case of "I want agentic coding without an API bill," the path exists and is documented.
"So it's like the difference between the genuine part and the pattern," Jake said. "The pattern works, costs less, and nobody's pretending it's the original." That's about the cleanest one-line summary of open-weight versus closed-weight that exists.
What Claude Opus 5.5 Actually Costs
Pricing is per million tokens — a token being roughly a small chunk of text, about three-quarters of a word in English, so a million tokens is around 750,000 words, several novels' worth. 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 staring at. A cached read happens when the model reuses a prompt prefix it has already processed — the front of a long conversation, a big document, your codebase context — instead of re-reading it from scratch. These make up the majority of cost in long agentic and coding sessions, and that line dropped 60%, not the standard 20% cut applied to input and output. For anyone running long coding sessions, the cache-read price is arguably the real headline number in this release.
Plain arithmetic to make it concrete: a heavy coding session that sends 2 million input tokens and receives half a million output tokens costs $8 + $10 = $18 at list price — before caching, which in a real agentic session typically pulls the effective cost well below that.
There's also a Fast mode: a high-speed configuration for Opus 5.5 running up to 2.5× faster, priced at $8 per million input and $40 per million output — double the standard rate for double the speed. It's available in research preview on Claude Code and on the Claude Platform, on consumption-based plans, or via usage credits on subscription plans. A real speed-for-price trade, not a free upgrade, and worth knowing exists before you complain that Opus 5.5 feels slow on a big task.
The deeper savings run past the sticker price: Opus 5.5 also tends to finish the same task using fewer tokens and fewer tool calls than Opus 5 did, and Anthropic claims this compounds with the price cuts to roughly a 40% total cost reduction on typical workloads — not just the 20% the per-token numbers suggest. A cheaper model that also needs less of itself to finish the job is where the real math lives.
How It Compares: Fable 5.1, GPT-6 Astra, and GPT-5.6 Sol
These are Anthropic's own launch numbers, with the honest caveat that always applies to vendor numbers on launch day:
| 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% | — |
(Elo, for the unfamiliar, is a rating system borrowed from chess where points measure relative performance — higher means the model won more judged comparisons, and gaps between scores are comparable within a benchmark.)
Two honest caveats belong right next to those numbers, both stated by Anthropic itself. 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: Opus 5.5's scores 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 instead — which Anthropic reports likely reduced Opus 5.5's own measured performance on those categories. In other words, the real ceiling on some of these numbers may be higher than shown, held down by the model's own safety behavior during the test itself.
Two real-world examples do more than the table. 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's 12, at 51% lower cost, but neither clearly won on output quality alone. And the example most relevant to anyone maintaining a legacy system: asked to audit a 200,000-line codebase, Opus 5.5 finished in under 3 hours where Opus 5 took over 20 — a gap large enough to change what's practical to attempt at all. The HAProxy translation shows the quality ceiling; this shows the throughput floor.
One difference you'll notice in the first conversation rather than any benchmark: the writing style changed. Opus 5.5 is less likely to use jargon and more likely to put the important information at the start of its messages, rather than burying the answer under preamble. If earlier Opus models felt verbose or hedge-heavy to you, this is the most immediately noticeable day-one difference.
How It Compares to Sonnet 5 and Haiku 5 (The Models You Can Use Today)
Opus 5.5 costs $4 in and $20 out per million tokens. Sonnet 5 costs $2 in and $10 out — exactly half. The question is never "which is best" (Opus, clearly) but when the quality gap justifies the price gap, and that's a workload question, not a spec-sheet question:
| Claude Opus 5.5 | Claude Sonnet 5 | Claude Haiku 5 | |
|---|---|---|---|
| API price (per 1M in / out) | $4 / $20 | $2 / $10 | Priced below Sonnet |
| Built for | The heaviest coding, reasoning, and long-running agentic work | Speed and cost balance | Fastest, lightest |
| Reach for it when | The task is complex enough that retries and failures are expensive | Most everyday work | High-volume, latency-sensitive work |
Here's the counterintuitive bit that makes Opus worth its price for hard tasks: a model that finishes a complex coding job in one pass with fewer tokens and fewer tool calls can cost less per completed task than a cheaper model that loops three times to get it right. That's exactly the efficiency argument behind Anthropic's 40%-real-world-savings claim. For a chatbot answering customer questions, Sonnet or Haiku is almost always the right call, and Opus tier is waste.
"So which one do I actually need for the shop?" Jake asked. "I'm not auditing 200,000 lines of anything."
"Then you don't need Opus, and that's not a compromise — it's the right tool," Ethan said. "Opus is the specialist you call when the job is hard enough that a wrong answer costs more than the token difference. Writing product descriptions, answering emails, summarizing reviews — that's mid-tier work all day. The mistake isn't using Sonnet. The mistake is paying Opus prices for Sonnet work, or burning a weekend on Sonnet retries for an Opus-sized problem."
✅ The one-line tier rule
If a failure would cost you more than the price difference between tiers, use Opus 5.5. If a failure means regenerating a paragraph, use Sonnet or Haiku and bank the difference. And hold your final judgment on the mid-tier until Sonnet 5.5 arrives in the coming weeks — it may move this line.
The Switch-from-ChatGPT Question, Answered Plainly
If your work is agentic coding, terminal-based tasks, or long-horizon knowledge work, Opus 5.5's lead over GPT-6 Astra on those benchmarks is real and consistent — and it costs less per completed task. If your work is conversation, creative writing, or multimodal input and output, the calculus is different, and GPT-6's broader multimodal capabilities may matter more to you than Opus's coding lead. The honest framing: Opus 5.5 is the coding-and-agents specialist; GPT-6 is the generalist. If you already pay for both, the question is which one you reach for first when a task is ambiguous — and no benchmark settles that for you.
The Safety Restrictions That Come With It
Opus 5.5 is the first Opus model to launch with the elevated safeguard tier Anthropic reserves for its most capable models, because its cybersecurity and biology capabilities are now comparable to the bigger Mythos-tier models. The practical consequences, in plain terms:
What changed between versions
- Before: Opus 5 shipped without the elevated safeguard tier — its cybersecurity and biology capabilities didn't trigger the same restrictions.
- Now: Opus 5.5 launches with safeguards similar to Fable 5.1's on cybersecurity, biology, and distillation resistance, as the first release under Anthropic's pace-the-frontier policy. Safety training was broadly similar to its predecessors, with alignment testing and pre-release evaluation by outside organizations.
- What that means for you: most cybersecurity tasks are automatically re-routed to Opus 4.8 behind the scenes; deep biology research is gated behind Anthropic's Life Sciences Verification Program for vetted organizations rather than open to everyone by default.
A few specifics worth knowing rather than discovering mid-project. Opus 5.5 launches with preserved thinking, an anti-distillation safeguard — distillation being the practice of using a big model's outputs to train a smaller copy of it, and preserved thinking blocks API users from editing Claude's prior reasoning context to extract it. It applies to API accounts created on or after August 31, 2026, so check your account creation date if you integrate programmatically. As covered above, thinking cannot be switched off on this model, at any effort level. It carries EU AI Act watermarking (as Fable 5.1 does), marking generated content as AI-produced where that compliance matters. And for the compliance buyers: zero data retention remains available — on Amazon Bedrock it's available with zero data retention by default, and offered the same way previous Opus models were on other platforms.
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 — live the same day.
- Google Cloud (Vertex AI) and Microsoft Foundry (Azure) — both confirmed at launch. Anthropic explicitly states Opus 5.5 is "now available on all platforms."
Since this site covers AWS regularly, the Bedrock path deserves the specifics. 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 matters when you're reconciling a cost report later. To use 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 use the Invoke and Converse APIs via the AWS CLI or SDK if that's what your existing code already speaks. - 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 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. Anyone selling you a "cheap Claude API" is reselling tokens with a margin, and you should read their terms before their price.
Update, October 7, 2026: OpenAI's GPT-6.1 Sol, released September 29, was benchmarked directly against Opus 5.5. On OpenAI's own tests it scored 2.2 points above Opus 5.5 on AutomationBench at about a third of the cost, and beat it on GDP.pdf at less than half the cost per task. It is also half the price per token, $2 and $10 per million against Opus 5.5's $4 and $20. These are one lab's chosen tests, so compare both on your own work. Our GPT-6.1 Sol guide has the full comparison with Opus 5.5 and Sonnet 5.5.
Update, October 8, 2026: The small model this page was waiting for arrived on October 7 as Claude Haiku 5.5, not "Haiku 5". It costs $0.10 per million input tokens and $0.50 output for prompts up to 100,000 tokens, a tenth of Haiku 4.5, with the same 1M context and 128K output as Opus 5.5 and adaptive thinking with an effort dial. Prompts over 100,000 tokens pay five times those rates, which no other current Claude model does.
The same announcement halved Sonnet 5.5 cache reads to $0.10 per million; Opus 5.5 cache reads stay at $0.20. If you are choosing a tier, Haiku 5.5 is now the worker under an Opus 5.5 planner rather than a compromise. Our Claude Haiku 5.5 guide has the full price table, the benchmarks next to Sonnet 5.5 and GPT-6 Luna, and the five request changes from Haiku 4.5.
FAQ
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 — Claude apps, Claude Code, the Claude Platform API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
What can I run locally instead of Opus 5.5?
An open-weight model, not a Claude. Tools like Ollama run open-weight models (GLM, DeepSeek, Gemma, and similar) on your own hardware, and recent versions speak the Anthropic Messages API — so agentic coding tools can be pointed at them instead of Anthropic's servers. You gain privacy and zero per-token cost; you give up the benchmark performance, the 1M-token context, and the safety training.
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 rates, with cache reads down 60%. Fast mode runs 2.5× faster at $8/$40. Anthropic claims real-world savings reach roughly 40% on typical workloads because the model also uses fewer tokens and tool calls per task.
Is Claude Opus 5.5 available on Amazon Bedrock?
Yes, confirmed live on launch day. It's billed through AWS Marketplace under the model provider rather than a generic Bedrock charge, with zero data retention available by default and geo-routing endpoints (us., eu., au., jp., global.) for data-residency needs.
How does Opus 5.5 compare to Claude Fable 5.1?
Anthropic claims it performs at roughly Fable 5.1's level on most everyday work, while beating it on several benchmarks (Terminal-Bench 4.0 at 66.4% vs 55.8%, GDPval-AA at 1846 vs 1735) and at meaningfully lower cost per task — Fable 5.1 is priced at $10/$50 per million tokens against Opus 5.5's $4/$20.
How does Opus 5.5 compare to GPT-6 Astra and GPT-5.6 Sol?
It beats both on most benchmarks Anthropic reports — Terminal-Bench 4.0, GDPval-AA, and Humanity's Last Exam — though GPT-6 Astra edges it out on AutomationBench (41.4% vs 40.0%). It's not a universal win across every measure, and all figures are vendor-reported.
Should I use Opus 5.5 or Sonnet 5 for my workload?
Rule of thumb: if a failure would cost you more than the tier price difference, use Opus 5.5. For everyday chat, drafting, and standard tasks, Sonnet 5 at $2/$10 is the right call. For complex one-shot coding where retries are expensive, Opus 5.5 can cost less per completed task despite the higher per-token price, because it finishes with fewer tokens and tool calls.
What's different between Opus 5.5 and Opus 5?
Lower pricing on every token type (especially cache reads, down 60%), faster output, fewer tokens and tool calls per task, a clearer writing style that leads with the important information, and a higher safety-safeguard tier matching Fable 5.1 on cybersecurity and biology. Thinking can no longer be switched off, unlike on Opus 5.
Can you turn off thinking mode in Opus 5.5?
No. Thinking is always on at every effort level on Opus 5.5 and Fable 5.1, including on the Claude API. On Opus 5, thinking could be turned off at lower effort levels — that toggle is gone on this model.
What are the effort levels in Opus 5.5?
A ladder of settings from low up to max, controlling how much internal reasoning the model does before answering. Low responds fast with minimal thinking; max allows extended reasoning on hard problems. One early report found max effort can exhaust the 128,000-token output budget on simple tasks without producing an answer — test before defaulting any production workflow to max.
What safety restrictions does Opus 5.5 have?
Cybersecurity tasks are largely re-routed automatically to Opus 4.8; advanced biology work is gated behind Anthropic's Life Sciences Verification Program; and preserved thinking (an anti-distillation safeguard) applies to API accounts created on or after August 31, 2026. It also carries EU AI Act watermarking, and zero data retention remains available.
What is Opus 5.5's context window?
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 — roughly 750,000 words of input in a single conversation.
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 available depending on data-residency needs. If your routing code pattern-matches model names, verify it distinguishes 5-5 from 5.
Will there be a Claude Sonnet 5.5 or Haiku 5.5?
Yes — 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. If you're choosing a mid-tier Claude plan today, it may be worth waiting for the mid-tier refresh.
Is Opus 5.5 good for coding?
It's Anthropic's strongest stated area — leading on Terminal-Bench 4.0 at 66.4%, with real-world examples including a 200,000-line codebase audit finishing in under 3 hours versus over 20 hours for Opus 5, and an HAProxy C-to-Rust translation passing nearly all regression tests at 51% lower cost than Fable 5.1.
Revision note. Updated September 23, 2026, one day after Claude Opus 5.5's launch, covering the release as announced September 22 across Claude apps, Claude Code, the Claude Platform API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Pricing, benchmark scores, safeguard behavior, and the Sonnet 5.5/Haiku 5.5 timeline are the parts most likely to be refined in the days after a launch — check Anthropic's pricing page before treating any number here as permanent, and re-check the effort and thinking behavior before wiring this into production. If you came here after being told "the new Claude is free and open," you weren't crazy for asking — that mix-up went around all week, and now you know exactly which parts are true, which aren't, and what to check before the next one arrives.