OpenAI just cut GPT-5.
GPT-5.6 Sol has been the go-to for reasoning-heavy tasks since it launched — code generation with complex constraints, multi-step analysis, the kind of prompts where you need the model to actually think instead of pattern-match. The problem was always the per-token bill. At the old rates, a single sophisticated agent loop with tool calls and reflection cycles could easily run $0.40-$0.60. Run that a few thousand times a day and you're looking at serious money.
The new pricing brings input tokens down to $2.40 per million (from $3.00) and output to $9.60 (from $12.00). Credits follow the same ratio. That's not a rounding error — it's a 20% reduction on the most expensive part of the stack.
What's interesting is the timing. This dropped right as Anthropic's Claude 3.5 Sonnet started gaining serious traction for the same workloads. Sonnet's pricing has been aggressive from day one, and their 200k context window made it a natural competitor for long-context reasoning tasks. OpenAI isn't dumb — they know developers were running side-by-side evals. This cut narrows the gap enough that switching costs become the deciding factor again.
For teams already invested in the OpenAI ecosystem — function calling patterns, assistant API threads, the whole tooling chain — this removes the financial pressure to migrate. The switching cost of rewriting prompt architectures and revalidating evals across providers is real. A 20% savings on the incumbent model often beats a 30% savings on a new one when you factor in engineering time.
The credit pricing adjustment is particularly relevant for enterprise accounts on committed spend. If you pre-purchased credits at the old rate, check your dashboard — some accounts are seeing retroactive adjustments applied to unused balances. Not universal yet, but worth opening a support ticket if you're sitting on a large credit pool.
One thing the pricing page doesn't highlight: the batch API discount still stacks on top. If you can tolerate async processing (and for eval pipelines, background analysis, nightly report generation — you usually can), you're looking at effectively 50% off the new rates. That puts GPT-5.6 Sol batch processing at $1.20/$4.80 per million tokens. For high-volume reasoning workloads, that's genuinely competitive with self-hosted 70B models once you factor in GPU costs and engineering overhead.
The real question is whether this signals a broader pricing strategy shift or a targeted defensive move. GPT-5.6 non-Sol pricing is unchanged. Embedding models unchanged. This feels surgical — protecting the reasoning tier specifically. Which makes sense: that's where the highest-value, highest-margin enterprise workloads live, and it's where competition is fiercest.
If you're building on Sol today, re-run your cost projections. If you evaluated it three months ago and walked away due to price, the math has shifted enough to warrant a second look. The model hasn't changed — but the economics just did.