Onset

OpenAI Cuts GPT-5.6 Sol’s Output Price by a Third

OpenAI Cuts GPT-5.6 Sol’s Output Price by a Third

The promotion runs through at least November 21 and puts Sol’s published token rates 20% below Anthropic’s Claude Opus 5.

Illustration by Onset News: cover image for OpenAI cuts GPT-5.6 Sol’s output price by a third
Illustration · Onset News

Latest update

Arena moves Sol onto two efficiency frontiers

  • Agent Arena: Code, 13.1% net improvement at $4.08 per task
  • Agent Arena: Work, 11.2% net improvement at $2.59 per task
View 1 earlier update

Sol’s price cut shows up at the task level

On DeepSWE v1.1, GPT-5.6 Sol at maximum effort scored 72.7% at a reported $6.47 per task. Claude Fable 5 scored 69.7% at $21.63 per task on the same 113-task, long-horizon engineering benchmark.

reach_vb @reach_vb

GPT-5.6 Sol Max was already the better deal on DeepSWE v1.1 - The recent price cut widens the gap even further. 🔥 Sol scores 72.7% at $6.47/task, compared with Fable 5 Max at 69.7% and $21.63/task. DeepSWE tests coding agents on 113 original, long-horizon engineering tasks.

That makes the earlier pricing story concrete: in this benchmark, Sol paired a slightly higher score with roughly 70% lower task cost. It does not prove Sol is cheaper for every workload, but it shows how the temporary token discount can widen an existing cost advantage in output-heavy coding-agent runs.

Full story

A developer asking GPT-5.6 Sol to generate one million tokens now pays $20 instead of $30. OpenAI put that 33% cut into effect August 21; one million input tokens, the prompt and other material sent to the model, now cost $4, down from $5.

Eligible ChatGPT Work and Codex customers who buy extra credits also get more use for their money. Pro, Plus and Business subscribers receive no larger included allowance, and their five-hour and weekly limits remain unchanged.

Token prices alone cannot show which model completes the same job more cheaply. But the promotion has OpenAI and Anthropic competing on the price of access as well as model performance.