Anthropic's Claude Fable 5.2 arrives and sweeps every board — while a wide gulf opens between the flagship models and the budget tier.
01 · The Bottom Line
Three things that matter this week
Anthropic takes the lead outright. Claude Fable 5.2, released Nov 9, tops both the Thinking Score (raw intelligence) and the Doing Score (getting real tasks done) — the first model this cycle to lead every component we track.
Google is a genuine #2, not a distant one. Gemini 4 Carbon actually wins the human-preference vote on Arena, even while trailing Fable on the technical benchmarks. The contest at the top is real.
The budget tier is a third of the frontier. The cheapest models (Haiku 5.5, Luna) score in the low 40s and 30s against a 100 leader — a reminder that "cheap and capable" still means a real trade-off.
How to read this brief. Every model gets three plain scores on a 0–100 scale: a Thinking Score (reasoning and knowledge, no tools), a Doing Score (agentic tasks — using tools to finish multi-step work), and an Overall Score that weights thinking at 65% and doing at 35%. 100 = the week's leader on that measure.
02 · The Leaderboard
Overall Score
Fable 5.2 sets the pace; Carbon and Astra form the chasing pack.
AnthropicGoogleOpenAIxAI solid = flagship · light = value tier
A two-tier field. Three flagships (Fable, Carbon, Astra) sit at 75 and above; everything else lands below 50.
Fable's one asterisk. Its Thinking Score is still missing one of three inputs (no public Arena vote yet, three days after launch) — but even a weak result there wouldn't move it off the top.
Two models can't be ranked yet. Gemini 4 Hydrogen (released today) and Grok 4.3 don't have enough published benchmark coverage for a full Overall Score — a data gap, not a verdict.
03 · Thinkers vs. Doers
Smart, or gets things done?
The two skills mostly track together — but not perfectly.
Above the dotted line = better at doing than thinking. Below it = stronger reasoning than hands-on execution. Most models hug the line.
Haiku flips the usual order. Among the value models it's the rare case that scores a touch higher on doing than thinking — useful if your work is task-execution, not trivia.
The frontier pulls away on both axes at once. Fable and Carbon don't trade one skill for the other; they lead on both.
04 · What Moved
The Starting Line
This is week one — the baseline every future issue measures against.
Two launches reset the board. Claude Fable 5.2 (Nov 9) and Gemini 4 Carbon (Nov 10) both entered as GA flagships this week and immediately took the top two spots.
A third launch landed today. Gemini 4 Hydrogen, Google's new value-tier model, shipped Nov 12 — too new for full benchmark coverage, but on the board from next week.
From next issue on, this section shows change. Who climbed, who slipped, and whether it came from thinking, doing, or both.
05 · The App Layer
It's not just the model — it's the app around it
The same model scores differently depending on the coding tool it's wrapped in.
AnthropicGoogleOpenAIxAI solid = flagship · light = value tier
Why this is a separate board. These scores measure a model plus its maker's coding app (Antigravity, Codex, Grok Build, Claude Code) — the "harness" that feeds it tools and context. The wrapper can swing results as much as the model does.
Google's app leads here. Antigravity CLI with Carbon tops the roster-matched configurations we can source this week.
Fable's app result is still pending. Claude Fable 5.2 launched after this index's last run, so its Claude Code score isn't published yet — expect it to reshuffle this board soon.
These are a different test suite and scale from the main Doing Score, so the numbers can't be compared one-to-one across the two boards.
06 · On the Horizon
What's coming
Confirmed
Claude Fable 5.2 — live Nov 9
Anthropic's new flagship, generally available. Its restricted sibling, Mythos 5.2, stays limited to approved organizations.
Confirmed
Gemini 4 Carbon — live Nov 10
Google's flagship, now open to developers, enterprise, and consumers.
Confirmed
Gemini 4 Hydrogen — live Nov 12
Google's fast, low-cost everyday model shipped today; full benchmark coverage expected next week.
Announced
Gemini pricing rises Dec 1
Carbon's standard API rate is scheduled to roughly double once the introductory window closes.
Rumored
GPT-6.2 and Grok 4.8 chatter
Unconfirmed reports point to near-term OpenAI and xAI updates. Nothing official, nothing on the board yet.
Question of the Week
If the best model and the best app built around it are no longer made by the same company — does "which AI is winning" even have one answer anymore?