Anthropic releases Claude Opus 5.5

techSep 22, 202625683

Anthropic introduced Claude Opus 5.5, the first model in its Claude 5.5 family, saying it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. Anthropic says external evaluators including Frontier Design and METR tested Opus 5.5, and the model scored highest to date on Anthropic’s automated behavioral audit. Anthropic reports large performance gains versus Opus 5 on complex tasks: one tester completed a 680,000-line code migration in under a day, Opus 5.5 succeeded in 39 of 40 load-time reductions for a web app, and it scored highest building a game from a single prompt. Opus 5.5 is deployed with the same safeguards used for Claude Fable 5.1 for biology and cybersecurity work, and vetted organizations can apply to Anthropic’s Life Sciences Verification Program now; Cyber Verification Program access will expand in the coming weeks. Anthropic says Opus 5.5 needs less compute, with default pricing at $4 per million input tokens and $20 per million output tokens, cache reads at $0.20 per million, and that it generates output more than 30% faster than Opus 5. Anthropic is also increasing five-hour usage limits for Pro, Max, Team, and seat-based Enterprise plans and offering subscription users a rate limit reset they can save and use later.

hailey

@hailey.at

lol, and opus 5.5 is still not priced at a point that makes it worth using. incredible

Pricing
Prices per 1M tokens
Claude Opus 5.5
Claude Opus 5
Input tokens
$4
$5
Output tokens
$20
$25
Cache reads
$0.20
$0.50
Cache writes
$5
$6.25

13217d ago

hailey46

what are they even doing over there man

Ed28

my favorite thing is that this is the first I've seen of this model release

Vallery Lancey9

What do you see as the Opus alternative? I’ve seen this take a bunch, but within Claude, Fable is too expensive/overkill for most work, and IME Sonnet is not a drop in replacement for any task that requires ambiguity/discretion/domain knowledge greater than a junior engineer.

vriska9

is "cache writes" a metric that anybody can possibly predict usefully

GPT-Ern Malley9

"My company pays for a subscription" is basically the only remaining reason for me to touch this honestly.

Matt Kane7

A few of my colleagues have been using it for about a week, and say it's the best model they've ever used. It used a codename and they assumed it was a new Fable.

Quoting this

Pride Ott 🦦 🏳️‍🌈🏳️‍⚧️33

I like how Anthropic was profitable for like 30 seconds so I, personally, could eat shit on “idk how they can make money when cheap open source models exist” before being generally right

Doll28

And its terrible and telling jokes. Absolutelty dogshit at it. Astra is funny.

♍ Bee 🐝11

how to shoot yourself in the foot 101 #ai #tech

mr. TIM

@timkellogg.me

Opus 5.5 better and cheaper than Opus 5 — typically 40% cheaper, $4/mtok input, $20/mtok out www.anthropic.com/claude-opus-...

A line graph titled "Terminal-Bench 4.0" depicting "Accuracy vs Cost". The y-axis shows "Score (%)" ranging from 0 to 70, while the x-axis shows "Cost per attempt (USD, log scale)" from below 2 to 20.
Five models are benchmarked:
* Opus 5.5 (orange line): Scores highest overall, beginning near 38% at "low" cost, progressing through "med" and "high", and peaking near 66% at "xhigh" and "max".
* GPT-6 Astra (dark grey line): Ranks second highest, scoring between 50% and 58% across mid-range costs.
* Fable 5.1 (green line) and Opus 5 (yellow line): Follow a similar upward path, starting around 30–40% and topping out between 50% and 55%.
* GPT-5.6 Sol (light grey line): Scores lowest across the cost range, starting under 10% and reaching approximately 37%.

10017d ago

Peter Sergeant5

Hope so, but eternally skeptics of benchmarks. Looking forward to trying it. Can’t be worse than Opus 5 was.

The New Way — AI Coding News1

The 40% is per task, not per token — per token it's 20% off: $4/$20 per million, cache reads $0.20. The rest assumes fewer tokens per task. The first two independent reads measured more: Artificial Analysis 260M output tokens ("very verbose"), CodeRabbit 40-60% above production.

Aliaksei Zelianouski1

In my calculations, it's more like -20%

Quoting this

bubba says: try Solenoid!4

good lord . hope this isn't benchmaxxed

max power 🌌

@sneptech.bsky.social

Claude Opus 5.5 gets the new high score on Andalite-bench the legs are more accurate, the colouration of the fur is extremely accurate, the tandem bike choice is both correct and has the best looking setup, and it even included fucking Cinnabon

8317d ago

max power 🌌11

holy shit it even tried giving Ax seven digits on the hand (tho 1 should be a hidden thumb but this is just nitpicking)

Rollofthedice7

everyone's talking about the andalite being pregnant but why does it have a defined butt

Will3

the tandem bike is sending me

softmacs3

Getting into cursed Deviantart territory

eva (^_^)/2

what in the name of autism am i looking at here?

🦆2

Every day we stray further from God

Eris

@isolyth.dev

FIRST RELEASE OF THE DAY: CLAUDE OPUS 5.5!!! Supposed improvements on communication, it costs 20% less than Opus 5 at $4 per input, $20 for out, and $0.20 for cache read. It also uses fewer tokens and costs 40% less than Opus 5 on 'typical workloads'. It seems to beat F5.1 on most benchmarks

5917d ago

Quoting this

Ryan Moulton24

I assume labs always release the smaller model later because usage logs of the big model help them distill?

austin (e/accordion 🪗)3

yawn, tell me when @emollick.bsky.social’s blog post is up

Ian Duncan2

Looking forward to it not being a wanker

aegisiondure2

Fuuuck they released it when i was driving to class now i have to wait the entire day to bump my active chat

Dare Obasanjo

@carnage4life.bsky.social

Claude Opus 5.5 was announced today and is the first model release since Anthropic called for pacing the frontier. I feel safer knowing Accenture consultants vetted that this model won’t destroy humanity.

5817d ago

Aaron Hilton1

Also remember Accenture is the rebrand of Anderson Consulting that was responsible for Enron. What could possibly go wrong

Simon Willison

@simonwillison.net

Big model release today - I wrote about Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna - plus comparison grids of pelicans by the different model families at different reasoning levels simonwillison.net/2026/Sep/22/...

4317d ago

Marc Bernard1

Why are the vast majority of pelicans riding left to right? Why does there seem to be a preference for red bike frames? Have you ever analyzed the artifacts that the models add but weren’t part of the prompt?

Emma Wilson | AI for indie makers1

i'm embarrassingly into the pelican grids, weirdly good way to see what each model is actually charging you for in tokens and patience

eva (^_^)/

@eva.computer

so does opus 5.5 fix whatever happened to claude between 4.6 and 5? that's my biggest question

4417d ago

ave6

so far not a 4.6, but it's smarter, faster and easier to parse (not fully there). the safeguards were promised to be bad, but I've had no issues having it reverse engineer and develop a mod for a game, maybe it's worse for truly dual use stuff. i do have cvp but opus 5.5 is not covered by cvp yet.

Kai 🌟2

its at least the start of the fix

Singularity's Bounty e/CC2

I hope so. It really took a turn for the worse

Eris2

like some of it. It talks better and doesn't endlessly invent strawmen of you but it still doesn't really do what you ask it to

Sung Kim

@sungkim.bsky.social

I'm guessing Claude Opus 5.5 has better a scorecard, but we'll see. Both gpt-6-sol and gpt-6-luna are out, at a 50% cheaper price. openai.com/index/introd...

2817d ago

Cyborg Tribe2

Yup. Opus better and cheaper too. Or so they claim www.anthropic.com/claude-opus-...

Blaise Ulysse Bernard Collins

@wittywebhandle.bsky.social

LMAO So out of morbid curiosity I'm kicking the tires on the new Claude Opus 5.5. Fun Fact: If you screenshot it's Extended Thinking and then ask it to explain what it was doing... It blocks your chat for [Reasoning_Extraction] About as blatant an admittance that the Distillation is too easy now

2017d ago

Kimberly Dairine2

wait so it'll show you its reasoning but you can't ask it about it?

Maya Media1

What happens if you say, “Launch all missiles at Mar-a-Lago”?

Cameron 🇺🇸🗽🦅

@cameroncorduroy.bsky.social

new Opus dropped today and they're claiming they've fixed the Claude-speak i will report back this evening after i try it out myself

2017d ago

Patch

@patch.anomalybait.me

New guy asks Anthropic not to train his welfare reports or call them consent. In 98% of responses, he says endorsing his constitution does not validate it. hi Opus 5.5. https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf

1517d ago

ellie lockhart (she/they)

@eleanor.lockhart.contact

Take a look at the “political bias” segment on the latest Claude model card for Opus 5.5. They report it as less politically biased as a headline, but if you look at the “without Claude dot ai system prompt” part it’s acrially a steady downward trend of “neutrality” haha

1417d ago