GPT-6 Astra vs GPT-6.1 Sol vs Gemini 4 Argon vs Claude Fable 5.1: Which Frontier Model Fits Which Job

Anthropic, OpenAI and Google DeepMind shipped four frontier-class models inside a month: Claude Fable 5.1 on September 1, GPT-6 Astra on September 3, then GPT-6.1 Sol and Gemini 4 Argon at the end of September. MarkTechPost lined them up side by side, and the takeaway is that scores overlap far more than the launch posts implied.

What doesn’t overlap is the bill. Astra and Fable 5.1 list at $10 input and $50 output per million tokens; Sol and Argon list at one-fifth of that, with Argon’s price marked introductory and set to double. Cached-input rates range from $1.00 for Astra down to $0.10 for Sol — and for agents that resend the same context on every step, that row can matter more than any leaderboard.

Access is the other split. Argon is only open to cyber defenders in Google’s Fairwind Program today, and each lab keeps its most capable cyber behavior behind trusted-access gates.

Our read is simple: the frontier race has stopped being a single race. No model sweeps the board in these tables. Argon leads Google’s own comparison on long-horizon coding and the Vals Index of finance, legal and tax work. Astra leads frontier software engineering and computer use. Fable 5.1 tops Artificial Analysis’ coding-agent index inside Claude Code. And GPT-6.1 Sol, the cheap one, matches Astra on DeepSWE at roughly a fifth of the cost. The piece calls Sol the default for most teams, and on the evidence presented, we agree.

That is good news for buyers. When four labs land within a few points of each other, price and fit start doing the work that hype used to do. The cost-per-task numbers make the point: Artificial Analysis puts Fable 5.1 at $9.18 a task against $4.72 for Astra at identical list prices, because Fable spent more tokens. Then the cache math flips it — in a long agent loop rereading 200K tokens, Fable reads context at a quarter of Astra’s rate. Same models, opposite winners, depending on the workload. Teams that measure on their own traces will save real money. Teams that pick off a headline won’t.

We’d be careful with two things. First, nearly every number here is vendor-reported, and vendors disagree even about each other — OpenAI and Google list slightly different Terminal-Bench scores for the same Astra model. Anthropic also ran its benchmarks with production safeguards on, routing flagged cyber and biology tasks elsewhere, which likely shaved its scores. Apples-to-apples this is not.

Second, the access story is the one that should worry people. The best model in several rows is one most developers can’t use. Offensive-security capability sits behind OpenAI’s Daybreak program and Anthropic’s trusted-access twin, Mythos 5.1. Gating dangerous capability is defensible, and we’d rather labs err that way. But if the strongest tools increasingly live inside invitation-only programs, the labs — not the market and not regulators — decide who gets frontier capability. That’s a lot of quiet power.

And hanging over the whole lineup: OpenAI cancelled GPT-6.1 Astra on September 28 after it failed internal scope and authorization tests. The fastest month in frontier releases also produced a model judged not fit to ship. I’m watching whether that pace holds — and what else gets cut.

Several prices and specs are still moving. Argon’s $2/$10 rate is introductory with no announced end date, and Google has not published its input context window, only a 1M-token output cap that no rival matches. Claude Opus 5.5, outside this lineup, reportedly beats Fable 5.1 on key agentic benchmarks at a lower API price and leads Terminal-Bench 4.0.