Claude Fable 5.2 drops to $8/$40, and Gemini 4 Carbon opens to developers at half price until December

The week on the beat

1. Anthropic released Claude Fable 5.2 at $8/$40 per million, down from $10/$50 for Fable 5.1

It's up to 30% faster, keeps the 1M context, and is on the API, Amazon Bedrock, Google Cloud and Microsoft Foundry now. Anthropic says it scores about level with Opus 5.5 at max effort for roughly half the cost per task. Fable 5.1 won't retire before September 2027, so you can switch on your own schedule.

Read the story →

2. Gemini 4 Carbon is now open to paid Gemini API developers, at $2/$10 until November 30

From December 1 it costs $4/$20, and so does Gemini 4 Argon. The free API tier doesn't include Carbon, and in the Gemini app and Antigravity it's rolling out to AI Ultra subscribers. Google says Carbon uses more tokens than Argon, so a task costs more even at the same rate. Budget for December's price now.

Read the story →

3. Fireworks AI disclosed a second breach, and Notion says AI chats from under 1% of its workspaces may have been exposed

An intruder got in through a third-party cloud provider on October 24 and reached billing data and, for some accounts, secrets and partial model weights and datasets. The first breach, on October 6, exposed API keys stored as evaluation-job secrets. Fireworks says it keeps no prompts by default, but Notion says the endpoints it used retained data. If you use Fireworks, rotate every credential its notice names.

Read the story →

4. Meta released Muse Ember, a small agent model at $0.10/$0.50 per million with a 1M context

That's Claude Haiku 5.5's per-token price, and Artificial Analysis scores both 36 on its index, but Ember costs $0.14 per index task against Haiku's $0.41. Its weak spot is knowledge: it declines 64% of the questions in Artificial Analysis's Omniscience test. It's on Meta's Model API as muse-ember-1.0, Google Cloud and Oracle Cloud. Use it for tool calls, routing and extraction, and keep lookups on a bigger model.

Read the story →

5. Google launched the Gemini Decisions API, which scores up to 1,000 yes/no, choice or rubric questions on one input in a single call

It returns calibrated probabilities instead of text, at $0.10 per million input tokens and nothing for output, with a free tier in AI Studio. Google pitches it for routing, tool selection and policy checks, at 150-250 ms a request. It's on the Gemini Developer API only, as gemini-decision-1. If you use a full LLM call as a classifier or router, benchmark this against it.

Read the story →

6. OpenAI released GPT-6 Cyber, a security model trained to stay inside the scope you set

It's limited to approved Daybreak Red customers through the API and Codex, with wider access "soon". OpenAI says it built working exploits for 948 Linux kernel vulnerabilities, more than five times GPT-5.6 Sol's 166 in the same harness, and found 81 high-severity Firefox flaws, many of them already fixed by Mozilla.

Read the story →

Model moves

• New: Claude Fable 5.2 (Anthropic). $8/$40 per 1M, $4/$20 on batch, and $16/$80 in fast mode, which Anthropic says generates up to 2.5x faster. 1M context, 128K output. Model page →
• Price rise on December 1: Gemini 4 Argon and Gemini 4 Carbon (Google). $2/$10 per 1M until November 30, then $4/$20. Batch and flex go from $1/$5 to $2/$10. Model page →
• New: Muse Ember (Meta). $0.10/$0.50 per 1M with $0.004 cached input, rising to $0.30/$1.50 for prompts over 200K tokens. 1M context. Model page →
• New: Gemini Omni 1.2 Flash (Google). Video generation at $0.07 a second for 720p, 30% below Omni 1.1, with clips up to 15 seconds, up from 10. On the Gemini API, Google Cloud, Flow and Vids. Model page →
• Follow-up: OpenAI image models. gpt-image-1.5, gpt-image-1-mini and chatgpt-image-latest leave the API on December 1. Model page →
• No other vendor list-price changes we could confirm on a lab's own page, and no benchmark revisions in the tracker this week.

Personal take

Most weeks I'll point you at a cheaper host for a model you're already using. This week it's the other side of that equation. Fireworks claims they don't store any prompts by default, yet Notion had to inform some workspaces that their AI chats may have been exposed because the endpoints Notion used did, in fact, store them. Before you send any user data to a cheaper host, get their retention policy in writing for the exact endpoint you're going to call, and make sure your secrets aren't going into anything they run for you. A lower price per token is of very little use if you can't be sure where those tokens end up.

Until next Thursday, Anmol