Frontier AI Models Compared: Fable 5.2 vs GPT-6, Gemini 4, Grok 4.7 and Open Weights

Claude Fable 5.2 now leads the independent leaderboard, but "best model" depends on whether you mean the highest score, the lowest bill, or the one you can actually call today. In the two weeks around November 9, 2026, Anthropic shipped Fable 5.2, Google announced Gemini 4 Carbon, Moonshot opened the weights of Kimi K3.1 Pro, and Tencent released Hy4. This post puts eleven frontier and near-frontier models in one frame: Anthropic, OpenAI, Google, SpaceXAI (Grok) and the open-weight labs.

We use independent scores first, lab-reported numbers second, then price, access and license. We flag every place where the numbers do not compare, because a surprising amount of cross-lab benchmarking is apples to oranges. For the Claude-only deep dive, read Fable 5.2 vs Opus 5.5.

TL;DR: the questions people ask first

The master table: eleven models, one view

Scores come from different sources, so the source is part of the number. "AA v5" is Artificial Analysis's Intelligence Index v5. Prices are list prices per million input and output tokens.

A podium of ranked AI models, illustrating the frontier model comparison between Fable 5.2 and rivals

Three footnotes keep this honest. The Fable 5.2 score of 65 comes from Artificial Analysis's page as we cited it on November 9, and we have not re-run it. Kimi K3.1 Pro was listed on v5 on October 27 as "weights not released"; the weights landed on November 4, so check the current label. Grok 4.7 does not appear in our v5 source table, so we show it without an independent score rather than guess.

How does Fable 5.2 stack up against GPT-6 Astra and Gemini 4 Carbon?

Anthropic's launch post is the only table that puts Fable 5.2, GPT-6 Astra and Gemini 4 Carbon on the same rows. It is Anthropic's table, so every row favors Anthropic's chosen benchmarks, but it is the best three-way data we have. Opus 5.5 is included as a reference.

Fable 5.2 wins eight of nine rows here. Carbon wins AutomationBench, and its Chartography score of 72.5% is within one point. Astra is second on FrontierSWE v2 and Terminal-Bench-Science. The pattern matches what independent boards showed for Carbon earlier: it is strong on broad composites and weaker where long tool-heavy agents are involved. Our Carbon vs Opus 5.5 vs Astra vs Grok 4.7 post has the independent view.

Two practical limits. First, Anthropic omits benchmarks it does not win, so absence is information. Second, Carbon is not callable by the public, which means a Fable 5.2 vs Carbon bake-off is a thought experiment for most readers today. See the Carbon launch coverage for how access is expanding.

What does each frontier model cost?

A bead diverted at a three-way pipe junction, a picture of routing work across frontier models by price

List prices span almost two orders of magnitude, from DeepSeek at $0.15 input to GPT-6 Astra at $10.

List price is not cost per task. In Artificial Analysis's v5 run, one Index task cost about $10.04 on Opus 5.5 (max), $7.76 on Carbon, $6.24 on GPT-6 Astra and $1.27 on GPT-6.1 Sol. Astra costs five times Sol per token and about five times per task ($6.24 against $1.27). We did not find a published Fable 5.2 per-task figure on that board, and Anthropic's own cost claims are matched-effort comparisons against Opus 5.5.

The practical rule is the same for every row: price a successful artifact, not a million tokens. Our Sol vs Astra cost note and the output-token inflation post explain how a cheaper token can still produce a bigger bill.

Where does GPT-6.1 Sol beat the premium tier?

GPT-6.1 Sol is the model to beat on value. OpenAI sells it at $2 / $10, one-fifth of Astra's standard rates, with a 95% cache-read discount ($0.10 per million cached input), and Artificial Analysis scores it at 55 versus Astra's 57. If your agent loop is cache-friendly and your tasks sit inside what Sol handles, that is a large saving for a two-point drop. The GPT-6.1 Sol launch coverage lists what to rerun before you move.

Fable 5.2 does not compete on price. It competes on the hardest long-horizon rows, where it leads Astra by 20 points on Remote Labor Index and 35 on RoboDojo. If your work does not look like those rows, Sol or Opus 5.5 is a saner default.

Is Grok 4.7 still a frontier model?

Grok 4.7 is the cheap live option, but it is not the leader. SpaceXAI shipped it on September 21, 2026 at $2 / $6, the same price as Grok 4.6, and our launch breakdown shows it winning EEBench and a legal-agent benchmark while trailing Fable 5.1 on CursorBench 4.0 and Terminal-Bench 4.0 (38.0% against 57.9%).

That Terminal-Bench 4.0 gap is the clearest cross-lab number. Opus 5.5 posted 66.4%, GPT-6 Sol 43% (an Artificial Analysis run) and Grok 4.7 38.0%; see our Grok 4.7 vs Opus 5.5 vs GPT-6 Sol comparison. Grok 4.7 is also on Amazon Bedrock. We did not find a Grok newer than 4.7 in released form as of this post, so treat Grok as the price floor among closed models, not the performance ceiling.

How close are open-weight models to the frontier?

Two rows of server blocks, one open and one closed, showing open-weight versus closed frontier models

Closer on some benchmarks, far on others. Three data points frame it.

Kimi K3.1 Pro (Moonshot). Released October 23 with weights on Hugging Face on November 4. It is a mixture-of-experts model with 2.8T total parameters, 104B active, 896 experts and a 1M-token context. Moonshot self-reports 65.7% on Terminal-Bench 4.0, against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra in Moonshot's own table. Artificial Analysis scored it 46 on v5, which is 12 points below Opus 5.5 and well above the other open weights. API price is $2 / $10. The license is MIT-style with conditions: a Model-as-a-Service business earning over $20 million in any 12 months must sign a separate agreement, and products with over 100 million monthly users or $20 million monthly revenue must show attribution. Full detail is in our Kimi K3.1 Pro post.

DeepSeek V4.1 Flash 1017. MIT-licensed, about 763B parameters, 1M context, and $0.15 / $0.60 off-peak. DeepSeek self-reports Terminal-Bench 4.0 rising from 31.2 to 42.3 in one point release, and Artificial Analysis scored it 32 on v5. That is slightly above Grok 4.7's 38.0% on the same benchmark at a fraction of the price. DeepSeek ran the Claude Code harness for that number, which matters, because harness choice moves agent scores. See the DeepSeek V4.1 Flash 1017 post.

Tencent Hy4 and others. Hy4 is a 770B MoE with vision and computer use, released November 4. Tencent's own tests rank it second to fourth of seven, and the final weights are planned but not released (our Hy4 post). A Qwen Code commit treats qwen4 as a 1M-token family, but the launch itself is unconfirmed (claimed vs verified). Smaller entrants such as Ripple-65B report strong self-run scores that nobody has reproduced yet (claimed vs verified).

The honest summary: on agentic coding as one vendor measures it, open weights are within one point of Opus 5.5. On the independent composite, they are 12 to 26 points below the leaders. Both statements are true, because the benchmark choice decides the answer.

Which model should I use for what?

Two dials for cost and power, a picture of choosing between frontier models by budget

A routing setup beats a single pick. A common pattern is a premium model for planning and a value model for execution: Fable 5.2 plans, Opus 5.5 or Sol executes, DeepSeek handles bulk classification. Our Fable 5.2 vs Opus 5.5 routing sketch shows one way to code it.

Why you cannot trust a single leaderboard

Four traps recur in cross-lab comparisons, and each shows up in the tables above.

1. Different harnesses. DeepSeek used the Claude Code harness for Terminal-Bench 4.0. Carbon ran in Antigravity CLI on Artificial Analysis. A model can rank first in one harness and second in another, as Carbon did against Opus 5.5.
2. Different effort settings. Fable 5.2 is reported at Xhigh or Max on some rows and Medium on others. Comparing a Max row against a competitor's default is not a fair test.
3. Different index versions. Kimi K3.1 Pro scored 55 on the Artificial Analysis index at launch on October 23 and 46 on v5 four days later. The rebuild changed the weights, so scores from different versions are not comparable.
4. Self-selected comparisons. Each lab picks its own rivals. Anthropic's table omits benchmarks it does not win. Moonshot's table compares against Opus 5.5 and Astra on its chosen rows.

Our guide on how to read AI benchmarks goes through these in more depth, and the Artificial Analysis v5 write-up explains why about 60% of that index uses private or held-out data.

Honest limitations

• We did not run these models. Every score is cited from the lab or an evaluator, and we have not reproduced any of them.
• Fable 5.2's 65 on Artificial Analysis is as reported in our November 9 coverage; index weights change between versions.
• Anthropic's three-way table is Anthropic's. Rows for GPT-6 Astra and Gemini 4 Carbon were run or reported by Anthropic, and one chart labels the Gemini model differently from the table.
• Gemini 4 Carbon is not publicly available. Its numbers cannot be checked by most readers.
• Missing data. Fable 5.2 has no published Terminal-Bench 4.0 or per-task cost on the board we cited, and Tencent Hy4 has no independent evaluation.
• Prices change quickly. Introductory rates, off-peak discounts and cache tiers all move the real bill.
• Open-weight license terms are summarized from the model cards. Read the license before you ship a commercial product.

What this means for what you build or pay

If you pay for frontier tokens today, the biggest lever is not picking the top-scoring model. It is routing. Most teams can send planning, search and ambiguous long runs to Fable 5.2, send routine coding to Opus 5.5 or GPT-6.1 Sol, and send bulk work to DeepSeek or a self-hosted Kimi, then measure cost per solved task.

If you care about data control, Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017 are the first open-weight options that agent builders can plausibly use in production, with the license and hardware caveats above. If you need the highest ceiling, Fable 5.2 leads on the numbers we can see. Run your own 20 to 50 tasks through two or three candidates before you commit.

Related on explainx.ai

• Claude Fable 5.2 vs Opus 5.5: benchmarks, cost and when to switch
• Claude Fable 5.2 and Mythos 5.2: pricing, migration and what changed
• Gemini 4 Carbon vs Opus 5.5 vs GPT-6 Astra vs Grok 4.7
• Gemini 4 Carbon launch: cyber defenders first
• GPT-6.1 Sol launch: pricing and benchmarks
• GPT-6 Astra launch: benchmarks and pricing
• Grok 4.7 launch: benchmarks and pricing
• Kimi K3.1 Pro: open weights, benchmarks and price
• DeepSeek V4.1 Flash 1017: open weights under MIT
• Artificial Analysis Intelligence Index v5

Primary sources: Anthropic's Fable 5.2 announcement, Moonshot's Kimi K3.1 Pro model card on Hugging Face, DeepSeek V4.1 Flash 1017 model card, Artificial Analysis Intelligence Index v5.

Scores, prices and access status are accurate as of November 10, 2026. Frontier model releases change weekly, so re-check each provider's pricing page before you build on this table.