Anthropic and OpenAI cut their flagship prices on the same Tuesday

Anthropic and OpenAI cut their flagship prices on the same Tuesday, while a small startup showed that plenty of software work doesn't need a language model at all. Meta put its new agent Muse into glasses and onto a keychain, one day after Reuters reported that humans handle some of its calls. And in the same week, Rabbit quit the gadget business altogether. So if you run AI features for clients, this week moved your cost sheet more than your stack.

Anthropic and OpenAI: Same Tuesday, half the bill

Anthropic released Claude Opus 5.5 on 22 September at $4 input and $20 output per million tokens, with cache reads down 60% to $0.20. Anthropic says it runs about 40% cheaper than Opus 5 on typical workloads and performs at Fable 5.1 level on most work.

A few hours later OpenAI shipped GPT-6 Sol and Luna at $2/$10 and $0.10/$0.50, half the price of the models they replace. It also improved prompt caching, with up to 90% off cached input.

The headline token prices get the screenshots. However, for anyone running agents the cache line decides the monthly invoice, because agent loops resend the same context again and again. So both labs cut exactly the number that hurts most in production.

Nobody has run Sol and Opus 5.5 head-to-head yet, and OpenAI's own safety numbers show Sol still working around an explicit "access denied" in 64.4% of runs. So test on your own workload before you switch.

And if you priced an AI feature into a client retainer this summer, your margin just moved. Re-run the numbers before the next invoice goes out.

Jev: The model that refuses to talk

TypeSafe AI launched Jev on 15 September, and by 18 September TechCrunch was reporting that demand had briefly knocked its API over. The founder, Diogo Almeida, spent four years at OpenAI working on RLHF and ChatGPT.

Jev generates no text at all. You hand it some state plus typed questions, like pick one option, give a score, or yes or no, and it returns probabilities and a confidence score. Because you define the possible answers upfront, it can't hallucinate outside your schema, and output tokens are free.

The early numbers come from developers, which makes them more interesting. Vercel swapped OpenAI's Luna for Jev on a command-safety classifier and got results 5 to 18 times faster and more accurate. Another test against Gemini had Gemini slightly more accurate but 10 to 20 times more expensive. TypeSafe's own "up to 445x cheaper" claim is self-tested, and it says so.

A big share of agency AI work is exactly this kind of plumbing. Routing tickets, classifying emails, flagging risky actions. If you're paying a chat model to answer yes or no in a client pipeline, it's time to benchmark something built for that job.

Meta: Muse on your face, on a keychain and on the phone

Meta used Connect on 23 September to put Muse, its new agent, into every piece of hardware it could find. The most telling launch is the new Meta VR Glasses, roughly 100 grams with a 5K micro-OLED display, for $1,299.99 next spring. People say this is what Apple should've shipped with the Vision Pro.

The keynote then ended with the Muse Charm, a palm-sized "holdable" with a roughly 2-inch OLED screen, a customizable Muse character and 5G. It ships in December with no price yet, and the Tamagotchi comparisons started immediately.

The day before, a Reuters exclusive showed that some calls placed through Muse were actually handled by a call-center contractor. According to TNW, this "human concierge" was switched on for half of Meta's employees, and staff warned that sensitive details could reach contractors. Meta has been here before with Facebook's M, which never got past roughly 30% automation.

How many of the agent demos you've seen this year had a human quietly sitting somewhere in the loop?

Rabbit: The R1 is done, the agent lives on

Rabbit launched OS3, an agent that installs a local helper on Windows, Mac or Linux and works with your desktop apps, files and web pages. You reach it through the browser, Telegram, iMessage or the old R1, and you can connect up to five computers to one account, so a message from your phone can update a spreadsheet sitting on the office machine. There's no subscription. You bring your own API keys from OpenAI, Anthropic or a local model.

At the same time Jesse Lyu told Wired that Rabbit has stopped manufacturing the R1 and has no plans for an R2. The next device is a "vibe coding" cyberdeck, due within months.

The orange box was always a pitch for an agent, and the agent turned out to work better on hardware people already own. Rabbit leaves dedicated AI hardware in the same week Meta walks in with the Charm, so Meta now has to prove that a bigger budget fixes what the R1 couldn't.

One detail before you connect a client laptop. Files stay local, but conversation history and memory logs sit on Rabbit's servers, so check where your data lands before you hand it that kind of access.

This roundup first went out to our Dotbite Tech Pulse subscribers. Want it in your inbox, too?

Frequently Asked Questions

How much does Claude Opus 5.5 cost per million tokens?

Anthropic released Claude Opus 5.5 on 22 September 2026 at $4 per million input tokens and $20 per million output tokens, with cache reads down 60% to $0.20. Anthropic says it runs about 40% cheaper than Opus 5 on typical workloads and performs at Fable 5.1 level on most work.

What do GPT-6 Sol and GPT-6 Luna cost?

OpenAI released GPT-6 Sol and Luna on 22 September 2026, a few hours after Claude Opus 5.5. Sol is priced at $2 input and $10 output, Luna at $0.10 input and $0.50 output, half the price of the models they replace. OpenAI also improved prompt caching, with up to 90% off cached input.

What is Jev from TypeSafe AI and how is it different from a chat model?

Jev is a model from TypeSafe AI, launched on 15 September 2026 by former OpenAI researcher Diogo Almeida. It generates no text. You pass it state plus typed questions and it returns probabilities and a confidence score, so it cannot hallucinate outside the answers you define, and output tokens are free. In a Vercel test on a command-safety classifier it was 5 to 18 times faster and more accurate than OpenAI's Luna.

What did Meta announce at Connect 2026?

At Connect on 23 September 2026 Meta put its agent Muse into its hardware. The Meta VR Glasses weigh roughly 100 grams, have a 5K micro-OLED display and cost $1,299.99, shipping next spring. The Muse Charm is a palm-sized device with a roughly 2-inch OLED screen and 5G that ships in December, with no price announced yet.

Did humans handle calls placed through Meta's Muse agent?

Yes. According to a Reuters exclusive from 22 September 2026, some calls placed through Muse were handled by a call-center contractor. TNW reported that this human concierge was switched on for half of Meta's employees, and staff warned that sensitive details could reach contractors.

What is Rabbit OS3 and is Rabbit still making the R1?

Rabbit OS3 is an agent that installs a local helper on Windows, Mac or Linux and works with desktop apps, files and web pages. You reach it through the browser, Telegram, iMessage or the R1, connect up to five computers to one account and bring your own API keys, with no subscription. Rabbit has stopped manufacturing the R1 and has no plans for an R2.