On August 21, OpenAI cut API pricing for GPT-5.6 Sol: input tokens drop from $5 per million to $4, output tokens from $30 to $20, and cached input from $0.50 to $0.40. The new pricing runs for three months, expiring on November 21.
The scope is spelled out precisely. API calls are billed at the new rates, and credits consumed inside ChatGPT Work and Codex drop along with them. Monthly subscription prices for the Pro, Plus, and Business tiers are unchanged. OpenAI's stated reason is that capability keeps advancing while efficiency keeps improving, and some of those savings are being passed on.
Running the numbers on real usage
OpenAI's own framing is "over 20%," but the discount isn't uniform across the three rates: input is down 20%, cached input is down 20%, and output is down by a third. Output took the deepest cut, so how much a given workload actually saves depends on its shape.
Take a coding agent as an example. That kind of task is input-heavy and output-light, and it re-reads the same code context repeatedly, so cache hit rates run high. Assume a single task consumes 1 million input tokens and 100,000 output tokens, with 80% of the input hitting cache. Under the old pricing: 0.2×$5 + 0.8×$0.5 + 0.1×$30 = $4.40. Under the new pricing: 0.2×$4 + 0.8×$0.4 + 0.1×$20 = $3.12 — a savings of about 29%.
Shift to workloads with a heavier output share, like chat or long-form generation, and the savings climb toward that one-third figure on output. Go the other way, to pure retrieval-style short Q&A where output runs only a few hundred tokens, and the savings land close to the 20% input discount. The same announcement can feel ten-plus percentage points different depending on the product it lands in.
Now sitting just below Opus 5
The cut moves Sol's spot on the price list. Anthropic prices Claude Fable 5 at $10 per million input tokens and $50 output; Opus 5 at $5 and $25. Sol's old rate of $5/$30 sat just above Opus 5; its new $4/$20 now falls below it.
The competition at this tier is direct. Developers willing to pay for a flagship model are typically the ones running it inside coding assistants, customer-support systems, or agent pipelines for long-running tasks — bills accumulate monthly, and switching is driven purely by unit cost. Anthropic's foothold in coding workflows and the steady undercutting from Chinese models at the same tier both put pressure on this pay-by-usage segment.
There's a second number easy to overlook: the absolute price of cached input. At $0.40, a cache hit now costs a tenth of an uncached token, and in long-context workloads with repeated calls, that ratio decides the monthly bill more than the headline rate does. Teams that bring their bills down are mostly the ones chasing cache-hit rate — list price is almost secondary.
Why only three months
Three months is a short window. The GPT-5.6 family launched as three tiers in early July, and by late July, Luna and Terra had already gone through a round of cuts — Luna down 80%, Terra down 20%. Now it's the flagship's turn, and OpenAI issued a time-limited price rather than revising the price list itself.
The upside of a time limit is an exit: on November 21 the price reverts automatically to $5/$30, with no need to announce a "price increase" or book the discount as a long-term commitment on the books. The cost is that developers know this too — it's an open question how much call logic they'll rewrite, or how much existing traffic they'll migrate, for a price with a countdown attached. The workloads most likely to actually move are the ones that were already shopping between providers.
"As we continue to push the frontier of capabilities while improving efficiency, we're dropping API and credit pricing of GPT-5.6 Sol by over 20% for the next 3 months."
Leaving subscription pricing untouched is telling in its own right. Monthly subscriptions are predictable, fixed revenue; usage is where the head-to-head competition plays out. This round of savings lands precisely on API and Codex credits — the accounts burning through thousands of dollars a month, the ones that can swap out a base_url at any time.
For developers in China, the real impact depends on which route they're using. Direct API calls get billed at the new price immediately; traffic through relays and proxies will have to wait for those providers to follow suit, and the gap in between tends to be exactly where the margin sits for the next few months. Once the price reverts in three months, whether that same traffic migrates back is a question that won't be answered until then.
Sources: OpenAI official announcement, CocoLoop, OpenAI Developer Community notice, Reuters; input, output, and cached-input rates and the three-month window cross-checked against OpenAI's pricing page.