Two days after OpenAI put it in developers’ hands, GPT-6 Astra took the one scoreboard that AI-assisted web developers actually argue about. On September 5, Arena.ai’s Code Arena: WebDev leaderboard seated GPT-6 Astra (Max) at the top with an Elo rating of 1,797 — 35 points ahead of Anthropic’s Claude Fable 5.1 (Max) at 1,762, and 109 clear of Claude Opus 5 (Max) at 1,688.
The margin is only half the story. The price tag is the other half. Arena.ai framed the result as a shift in the cost-performance curve, not just a rank change: GPT-6 Astra is, in its words, “SOTA performance at $40/Mtoken, matching the latest Claude model pricing.” Both flagships list at $10 per million input tokens and $50 per million output tokens; the $40 figure Arena plots on its Pareto chart is an output-weighted blend of those two rails. Crypto Briefing, covering the same board, described Astra as “putting 35 points of daylight” between itself and Fable 5.1.
For the first time in months, the frontier is not a question of paying more for a better model. It is a question of which identically priced model you point at which job.
A four-day launch window
Anthropic moved first. Claude Fable 5.1 and its trusted-access twin Mythos 5.1 went generally available on September 1, holding the same $10 / $50 sticker as Fable 5 while cutting cache reads from $1 to $0.25 per million — a 4x reduction. Anthropic estimates that math makes typical workloads roughly 25% cheaper and highly agentic work about 45% cheaper. The model carries a 1,048,576-token context window, 128K max output, always-on adaptive thinking, and a June 2026 knowledge cutoff. Its self-reported launch table, run with production safeguards on, shows Terminal-Bench 4.0 at 55.8 (against Fable 5’s 42.0), Terminal-Bench-Science 0.1 at 52.6 (against 24.7), CursorBench 3.2.0 at 73.4, OSWorld 2.0 at 41.7 strict, and GDPval-AA v2 at 1853 Elo.
OpenAI announced GPT-6 Astra on September 3 and released the `gpt-6-astra` API model on September 4. It ships with a 1,050,000-token context window, 128K max output, an April 30, 2026 cutoff, and two new rungs on the reasoning ladder — `xhigh` and `max` — on top of low, medium, and high. Cached input runs $1 per million, cache writes $12.50, and prompts above 272K input tokens bill at 2x input and 1.5x output for the entire request. OpenAI’s self-reported numbers: Terminal-Bench 4.0 at 57.7, DeepSWE v1.1 at 74.1, Terminal-Bench-Science 0.1 at 64.6 against Fable 5.1’s 52.6, GPQA Diamond at 96.0, BrowseComp at 91.5, and ARC-AGI-3 at 99.9.
Analysts have been careful to label the provenance of those figures. “Astra is the GPT-6 flagship, not a Flash hop,” wrote Sebastian Crossa, co-founder of LLM Stats, in his launch teardown — before adding the caveat that matters: “Hold those numbers as theirs, not as LLM Stats verified.” OpenAI president Greg Brockman was less measured, calling the model a “generational leap in capability.”
Independent scoring is genuinely split. On Artificial Analysis’ Intelligence Index v4.2, Fable 5.1 scores 57 at maximum effort to Astra’s 55, and leads the Coding Agent Index 70 to 67. Arena’s WebDev board says the opposite. Both can be true, because they measure different things.
Why It Matters
Code Arena is not a static benchmark, and that is precisely why this result landed harder than another vendor slide. It is a crowdsourced, Elo-style rating built from human preference votes — more than 650,000 of them across 126 models — where users compare paired outputs from models that had to plan with tools, build a live web app, and iterate. In a week when both labs published their own benchmark tables with their own harnesses and their own safeguard configurations, a third-party vote count carries weight self-reported rows cannot.
But the price parity is the load-bearing detail, and the sticker hides the real economics. At matched $10 / $50 list rates, cache is where the bill actually diverges: Fable 5.1 reads cache at $0.25 per million, Astra at $1. For long-running agent loops that replay the same context hundreds of times — the dominant shape of production coding work — that 4x spread can swamp a 35-point Elo gap. Astra’s long-context multiplier above 272K tokens cuts the same way for teams stuffing entire monorepos into a prompt.
The practical read for engineering leads is not allegiance but routing. Astra takes greenfield web-app generation, hard computer-use tasks, and research-grade science, where its Terminal-Bench-Science edge of 64.6 to 52.6 is the widest official gap on either company’s page. Fable 5.1 takes cache-heavy agentic loops and the sustained knowledge work where its cache pricing and Artificial Analysis lead compound over a billing cycle. Meanwhile, Alibaba’s Qwen 3.8 Max 0902 sits at 1,686 on the same board at roughly $5 per million — 111 points back for an eighth of the blended price.
What to watch
Three things will settle whether this is a durable shift or a launch-week spike. First, verification: LLM Stats has not independently confirmed Terminal-Bench 4.0, DeepSWE, or OSWorld for either model, and Terminal-Bench-Science carries roughly four points of standard error. Second, vote volume: a 35-point lead on a board this young can compress, and Arena is still collecting scores for Astra in its long-horizon Agent Arena. Third, cache economics: if Anthropic extends the $0.25 rate across the Claude family, or ships Enterprise Frontier Safeguards this fall as promised, the price story changes again — with an Anthropic IPO reportedly looming, neither lab is done repricing the frontier.
“Astra is the GPT-6 flagship, not a Flash hop. Hold those numbers as theirs, not as LLM Stats verified.”— Sebastian Crossa, Co-Founder, LLM Stats