GPT-6 Astra hits the API and ChatGPT Work gains a Data Agent while traders bet on Anthropic

What OpenAI shipped

OpenAI's busiest release stretch of 2026, as recounted in a dev.to analysis, centered on two products. GPT-6 Astra reached the API on September 3, following what the post describes as the company's largest training run to date, using more than 100,000 GPUs at the Stargate site in Texas. Input pricing sits at $10 per million tokens, which the post says matches Claude Fable 5.1 exactly.

A week later, on September 10, ChatGPT Work gained a Data Agent that answers natural-language analytics questions across Snowflake, BigQuery, and Redshift connectors. One detail the post argues most coverage missed: the agent is built on GPT-5.6 rather than Astra. Sam Altman has also teased further announcements for DevDay on September 29, after a promised mid-September launch slipped by a week.

Traders kept backing Anthropic

The post's central observation is that Polymarket's "best AI model" markets barely moved despite the launch blitz. As of September 18, it reports Anthropic priced at 98 percent for September against OpenAI's 1 percent, 88 percent versus 2 percent for October, and 72 percent versus 10 percent for year-end 2026. Anthropic's October odds actually rose 16.5 percentage points during launch week. The post pegs total liquidity across the three horizons at roughly $3.1 million, with about $51,000 in daily volume, and counts this as the third consecutive OpenAI flagship launch, after GPT-5.6 Sol in July, that failed to shift the odds.

Where the benchmarks land

Per the post, GPT-6 Astra scores 61 on the Artificial Analysis Intelligence Index while Claude Fable 5.1 scores 66. On Arena's most-watched text and overall categories, Astra carries no rating at all, while Fable 5.1 leads at 1,231 points and Claude Opus 5 Max tops the overall leaderboard at 1,505 Elo. Astra does lead WebDev Arena at 1,797 points, and the post notes that the October WebDev market is the only Polymarket AI market where OpenAI is ahead, at 62 percent.

There is also a cost angle: the post reports that Astra matches Fable 5.1's coding-agent performance at roughly 60 percent of the cost per task, a real advantage for high-volume workloads, though prediction markets track "best" rather than "cheapest."

How developers reacted

According to the post, Hacker News discussion of the Astra launch focused on pricing and cybersecurity rather than any dethroning of Claude, and a harness-design study trending on the site with 178 points argued that the infrastructure around a model matters as much as the model itself. The analysis also points to OpenAI co-founder Andrej Karpathy describing Anthropic's developer experience as a paradigm shift, and cites Ramp card-spend data from May showing enterprise spending on Anthropic at 34.4 percent versus OpenAI's 32.3 percent, with Anthropic adoption growing fourfold year over year.

Why it matters

The dev.to piece argues that builders should stop treating launch announcements as signal and lean instead on blind-evaluation leaderboards and enterprise benchmarks such as SWE-Bench, LiveBench, and the Intelligence Index. On that reading, an unmoved prediction market during a launch week is itself information, and Astra's lower cost per coding task is the practical reason a developer might still pick it. The post concedes its own bear case, though: Polymarket measures which model wins blind tests, while OpenAI's Data Agents, ChatGPT Work integrations, and roughly 400 million monthly users are distribution plays a leaderboard cannot capture. With DevDay set for September 29, developers will soon get another data point on whether this competition is about the best model or the best platform. Note that this story rests on a single commentary post, so its market and benchmark figures are its own reporting rather than independently verified numbers.