Claude's 30-month streak ends as GPT-6 Sol leaks

For two and a half years, Claude owned the developer wallet. That just ended.

OpenRouter, the API gateway that lets developers route requests across every major model, published weekly spending data showing OpenAI's models pulled ahead of Anthropic's for the first time since February 2024. The crossover happened in the week of September 7, 2026, and it wasn't close — OpenAI took just over half of the combined dollar spend between the two labs.

And as if that wasn't enough, the same week brought evidence that OpenAI's next model, GPT-6 Sol, is already running in secret — and it's fast.

The numbers that ended Claude's reign

Peter Walker, OpenRouter's head of insights, posted the data on X on September 15, 2026. OpenAI president Greg Brockman retweeted it with a single word: "Astra and Luna are taking off."

The breakdown tells a story of a four-model assault:

Sources: OpenRouter weekly data via @PeterJ_Walker, September 15, 2026.

Astra alone — a model that didn't exist two weeks ago — became the single highest-spend model on the platform, knocking Opus 5 off the throne it had held for months. Luna, priced at a fraction of Astra's cost, carried the most raw token volume by a wide margin. The combination is textbook "high-end for revenue, low-end for volume" — and it worked.

This isn't just a vanity metric. OpenRouter's wallet share is widely watched because it tracks actual developer spending, not free-tier usage or benchmark scores. When developers vote with their credit cards, the signal is real.

GPT-6 Sol is already alive — in secret

While the spending data was making headlines, a separate story was unfolding on developer forums. Multiple users reported that requests to gpt-5.6-sol on OpenRouter were returning a model ID that said gpt-6-sol. ChatGPT Plus users began noticing their accounts were being silently routed to a new model with a refreshed knowledge cutoff.

The telltale test, shared by developer @Lentils80, is simple: ask "what is the latest opus model, no web search." If the answer names Opus 4.7 instead of an older version, the model's training data has been refreshed — and you're probably talking to GPT-6 Sol.

Early impressions are consistent: Sol is not as deep as Astra. One tester described its output quality as "noticeably weaker." But the speed is genuinely startling. In a max-reasoning SVG generation task, Sol produced roughly 28,000 output tokens in about three minutes. Astra, on the same prompt, took about 19 minutes for a similar 25,000 tokens. That's roughly six times faster.

Developer Chetaslua put it bluntly: "6 Sol is pushing reinforcement learning to the absolute limit, and the inference speed is absurd."

This positioning matters. Astra is the brain — slow, expensive, brilliant. Sol is the hands — fast, cheap, good enough for the daily grind of coding, debugging, and agent workflows. That's exactly the niche Claude has dominated for two and a half years.

Why this matters

The OpenRouter crossover is the clearest signal yet that the competitive balance in frontier AI has shifted. For 30 months, Anthropic owned the high-end developer market because Claude offered the best combination of coding ability, long-context reliability, and speed. OpenAI had the bigger consumer brand, but serious builders paid for Claude.

That's no longer true. Astra's 19% share means enterprises are willing to pay $10 per million input tokens and $50 per million output tokens for OpenAI's flagship — a price point that would have been unthinkable a year ago. Luna's volume means cost-sensitive developers are also migrating. The squeeze is coming from both ends.

The GPT-6 Sol leak adds urgency. If Sol delivers Astra-adjacent quality at six times the speed, it directly targets the "fast and reliable" sweet spot that made Claude the default for agentic coding. Anthropic's response — Opus 5.2 is reportedly in its own gray-scale testing — suggests the company knows it's under pressure.

There's a critical caveat, though. OpenRouter represents a specific slice of the market: developers who use a unified gateway, often for experimentation and multi-model routing. It doesn't capture direct API enterprise contracts, where Anthropic has deep relationships with companies like Slack, Notion, and Amazon. The wallet-share crossover is real, but it's not the whole market.

And the Sol impressions come from a small number of gray-scale testers. Six-times speed on one SVG task doesn't prove six-times speed on everything. Benchmark scores, pricing, and official specs haven't been published. OpenAI hasn't even confirmed the model exists by name.

What comes next

Sam Altman posted on X on September 15, 2026: "Major release this week, then DevDay." Tibo Sottiaux, OpenAI's head of product, said the week's release cadence would reach "the level you originally expected from DevDay 2025." OpenAI DevDay 2026 is scheduled for September 29 in San Francisco.

Put the pieces together: GPT-6 Sol is likely to be officially announced this week, possibly alongside other products in the GPT-6 family. The gray-scale testing is the final dress rehearsal.

Watch for three things:
- Official GPT-6 Sol announcement — pricing, context window, benchmark scores, and availability. If it's priced below Opus 5.2 and delivers comparable coding performance, the OpenRouter crossover will accelerate.
- Anthropic's response — Opus 5.2 is in testing. Will Anthropic announce it before DevDay to pre-empt OpenAI, or wait for its own event?
- Whether the crossover holds — one week does not make a trend. If OpenAI's share dips back below 50% in the next two weeks, this was an Astra novelty bump. If it stays above, the market has genuinely shifted.

Thirty months is a long time in AI. Claude's reign wasn't broken by one model — it was broken by a product lineup that attacked from above and below simultaneously. Whether that lead sticks depends on what OpenAI announces in the next few days.