MODEL

GPT-6 Sol

modeltopic-noteopenai

Overview

GPT-6 Sol is OpenAI‘s Sept 22, 2026 coding-tuned frontier model, extending the GPT-6 generation (which shipped Astra earlier this month). Sol prices at $2/$10 per Mtok input/output — down from GPT-5.6 Sol at $4/$20, roughly half — and OpenAI claims Sol makes “about half as many mistakes” as GPT-5.6 Sol. Lands in ChatGPT Work, Codex, and the API for paid tiers, shipped alongside GPT-6 Luna and within a 12-hour window of Anthropic‘s Claude Opus 5.5 release — the sub-24h paired frontier launch with matched per-token cuts is what’s genuinely new.

Timeline

  • 2026-09-23-AI-Digest — OpenAI extends the GPT-6 generation with GPT-6 Sol on Sept 22 — coding-tuned, priced at $2/$10 per Mtok input/output (down from GPT-5.6 Sol at $4/$20), and OpenAI claims Sol makes “about half as many mistakes” as GPT-5.6 Sol. Sol lands in ChatGPT Work, Codex, and the API for paid tiers. The 12-hour overlap with the Claude Opus 5.5 release is what’s genuinely new — Simon Willison flagged the sub-24h dual launch with matched per-token cuts as unprecedented; the pricing move sits on an 18-month downward curve, a rung not a war ignition, but “both leaders cut on the same day” is the coordinating datapoint the market will re-price against.

  • 2026-09-24-AI-Digest — Referenced as reasoning-tier anchor in the “Tokens too cheap to meter” HN discussion (269 pts · 188 cmts). The “cheap enough to stop metering” thesis holds cleanly on the commodity tier (GPT-6 Luna at $0.10/$0.50, Gemini 3.8 Flash TTS at $9/Mtok intro) but breaks at reasoning/frontier — gpt-5 (high) reasoning tokens and Claude Opus 5.5 at $4/$20 remain expensive enough that per-call accounting is worth its overhead. Read tier-specifically. Also: Aider polyglot top-5 rescore hasn’t rotated to include Sol yet — measurement lag continues into the third day post-launch.

  • 2026-10-19-AI-Digest — GPT-6 Sol tops the Microsoft + CityU HK Taste-Bench panel at 59.7% on the top tier of 502 decision-fork questions that test whether an agent picks the better of two trajectory continuations given identical visible state. Across the full 14-model panel the average collapses from 62.3% → ~21% when the deciding clue only surfaces after extended downstream work, and reasoning-budget scaling barely moves the number. Load-bearing softener: 59.7% is the top-tier number on a specific decision-fork benchmark, not a general frontier-leader claim — the real benchmark signal is the long-horizon collapse (62.3% → ~21%), which hits all panel models including Sol. The practitioner result ships with a usable training signal: a distilled Qwen3.6-27B mimicking Sol-adjacent teacher-judge output doubled SWE-bench Pro success from 14.6% → 33.7% on 41 held-out tasks when bolted on as an advice channel. Reframe worth carrying: agent-eval cluster, not an agent self-improvement recipe — Taste-Bench is one measurable intervention among three (CarryOn’s 51.9% token cut, Self-HPO’s 57.8% → 64.8% resolve-rate lift on Qwen3.5-35B-A3B) hitting different failure modes. Log against MOC - Agentic Coding.

Key Developments

  1. Priced at $2/$10, Roughly Half of GPT-5.6 Sol (September 22, 2026): Standard pricing at $2/$10 per Mtok input/output, down from GPT-5.6 Sol‘s $4/$20 — the underlying vector is coding-tune plus a per-token cut on the frontier tier. OpenAI’s “about half as many mistakes” claim is vendor-cited; independent replication is the standard corpus caveat.

  2. Sub-24h Paired Launch with Claude Opus 5.5 (September 22, 2026): The 12-hour overlap with Anthropic‘s Claude Opus 5.5 release is the load-bearing distinguishing feature — Simon Willison flagged the sub-24h dual launch as unprecedented. Near-simultaneous releases themselves are a recurring pattern (Fable 5 vs GPT-5.6 preview in June was similar); the wrinkle today is paired price cuts dropping the same day. Framing to carry: coordinating datapoint on the price curve, not price war ignition.

  3. Ships in ChatGPT Work, Codex, and the API for Paid Tiers (September 22, 2026): Sol is the paid-tier coding-agent flagship inside the split with GPT-6 Luna — Luna reaches free ChatGPT users while Sol carries the reasoning premium. Tier-stratification hardens the pattern GPT-5.6 Sol / GPT-5.6 Luna established in July.

See also: OpenAI, GPT-6, GPT-5.6 Sol, GPT-6 Luna, Astra, Claude Opus 5.5, Simon Willison, MOC - Major Companies, MOC - Agentic Coding.