Gemini 4 Carbon

On November 10th, Google opened Gemini 4 Carbon to everyone who pays: the paid Gemini API, Gemini for business, and AI Ultra subscribers. Since October 28th it had been limited to selected security teams.

The price is the story. Two dollars in, ten out, per million tokens, until November 30th. Then it doubles, to four and twenty. Claude Fable 5.2 is eight and forty.

The catches. On FrontierCode 2.0, Fable leads by eight points, 49 to 41. Google also claims first place on WANDR, and Anthropic’s table leaves Carbon’s cell blank, so that stays a claim. And the launch price is temporary. Budget on four and twenty.

Fable 5.2

Anthropic shipped Claude Fable 5.2 on November 9th, with Mythos 5.2: the same model with fewer safeguards, for vetted security and life-science teams. The change you can verify yourself is the price. Eight dollars in, forty out. Fable 5.1 was ten and fifty.

The benchmark table is Anthropic’s own: 69.9 percent on FrontierSWE version 2, against 65.5 for GPT-6 Astra, with no outside evaluator named. Treat it as a claim.

Independently, it tops the Artificial Analysis index at 65. That is at max effort. At high effort it scores 62, and time to first answer drops from about four minutes to about half a minute. Cursor’s table shows the same shape: max buys three points over high, for more than three times the cost per task.

Two things in the fine print. Flagged cyber and biology queries get routed to a less capable model, though Anthropic says about 55 percent fewer get blocked than on 5.1. And Fable requires 30-day data retention by default.

Muse Ember

The same day, Meta released Muse Ember, its smallest model, built for the repetitive parts of agent work. Ten cents in, fifty cents out, per million tokens, with a one-million-token context.

Now the catches. The chart’s caption calls Ember’s costs estimates at proposed pricing. Four of the benchmarks on the page are Meta’s internal evaluations. On FrontierCode 2.0, run by Artificial Analysis, Ember loses to Haiku: 27.2 against 34.9. And past 200 thousand input tokens, the price triples.

Gemini Omni 1.2 Flash

On November 5th, Google made its video model, Gemini Omni 1.2 Flash, generally available on the Gemini API and, for the first time, on Google Cloud. Clips go up to 15 seconds, from 10. The price is 7 cents per second at 720p, which Google says is 30 percent below Omni 1.1. A full 15-second clip is a dollar five. You can also fix a shot with a follow-up edit instead of regenerating it.

What’s missing: the announcement has no benchmarks and no third-party comparison, only Google’s own side-by-sides. Google does name what it still gets wrong: fast sports, rapid script changes, complex physics.

GPT Image 2.6

On November 4th, OpenAI released GPT Image 2.6 as two API models, Flare and Sunburst. New in both: SVG output and animated GIFs, straight from the image API.

The rate card didn’t move: eight dollars in, thirty out, per million image tokens, identical to 2.5.

Your bill can still move, because you pay per token and 2.6 uses more of them. One independent test, 46 images at one size: at low quality, 2.6 Flare used 664 output tokens where 2.5 used 158. That’s 2 cents an image against half a cent. Four times more. At medium, about three times more. At high, the two cost the same. And in that test, auto picked low, so the default request is the one that got four times dearer.

OpenAI’s own guide says equal token rates don’t mean equal cost per image. And there is no batch price for 2.6 yet.