Bottom line up front: A new mid-tier OpenAI model is expected this week at roughly one-third of current flagship pricing. The AI model price war is compressing in UK businesses' favour — but you need a ready test case to evaluate it quickly, not a reactive scramble after the launch announcement.
What GPT-5.6 is
GPT-5.6 is expected to be a mid-tier model positioned between current GPT-5 and GPT-5.5 in terms of capability, with particular improvements in reasoning, coding, and vision tasks. The internal codename "kindle-alpha" has appeared in OpenAI's Codex backend logs since mid-May 2026, suggesting the model has been in internal testing for several weeks ahead of a public release.
At time of writing, OpenAI has made no official public announcement or confirmation of a launch date. The 83% probability figure comes from prediction market Polymarket, where traders with financial stakes are pricing the odds of a GPT-5.6 launch by 28 June 2026. Polymarket predictions have tracked OpenAI launch timelines reasonably closely over the past 18 months, though they are market signals rather than guarantees.
If GPT-5.5 is currently priced at approximately £5 per million input tokens at the API level, GPT-5.6 is expected at approximately £1.60–1.70 per million input tokens. Output token pricing follows a similar ratio. This is an estimate based on OpenAI's historical mid-tier model pricing patterns — actual pricing will not be confirmed until the launch announcement.
When it is expected
The current Polymarket odds give 83% probability of a GPT-5.6 launch by 28 June 2026 — within the next six days at time of writing. That said, OpenAI has slipped model launches before without warning, and there is no official confirmation to anchor the expectation firmly. The signal is strong enough to prepare for this week; it is not strong enough to plan workflows that depend on it launching on a specific day.
The most practical stance is to prepare your evaluation now so you can run it within 48 hours of whenever the model does land, rather than scrambling to evaluate it reactively.
Why it matters for UK businesses
The AI model price war between OpenAI, Anthropic, and Google is one of the most unambiguously positive trends for UK SMBs in the current market. Each new model generation has tended to deliver better performance at lower cost than its predecessor. GPT-5.6 appears set to continue that pattern.
GPT-4 (March 2023): ~£24 per million input tokens. GPT-4 Turbo (November 2023): ~£8 per million input tokens. GPT-4o (May 2024): ~£4 per million input tokens. GPT-5 (2025): ~£9 per million input tokens. GPT-5.5 (early 2026): ~£5 per million input tokens. GPT-5.6 (expected June 2026): ~£1.60 per million input tokens. The direction is consistent: more capable models at lower prices, quarter after quarter.
For UK SMBs, a model that costs £1.60 per million input tokens instead of £5.00 changes the economics of document processing, customer communication drafting, data extraction, and analysis workflows. At that price point, tasks that were marginal cases for AI automation — too expensive to justify at scale — become straightforward business cases.
The practical implication is not to switch platforms mid-workflow based on a model announcement. It is to evaluate deliberately: test GPT-5.6 on your actual use case, measure the quality against what you currently use, calculate whether the price-performance improvement changes your automation economics, and then decide with evidence rather than headlines.
What to do
One action: write your GPT-5.6 test case now, before the launch
Identify the specific task you want to evaluate — a proposal draft, a customer email sequence, a data analysis job, a coding function. Write down exactly what "better than your current model" looks like for that task: output quality, accuracy, format, tone. Define your success criteria before you see the model, not after. This prevents the cognitive bias of evaluating a new model against a vague standard and concluding it is good because it is new. When GPT-5.6 launches — this week or next — run the prepared test within 48 hours. Compare output quality against your current tool on the same task. Factor in the pricing difference. Then make a deliberate decision: switch, keep testing, or stay with your current setup. Do not defer the evaluation to "when things are less busy". Deferred evaluations become permanent non-decisions.
The broader point is that competitive AI model evaluation is now an ongoing business skill, not a one-time IT decision. The market moves fast enough that a model you evaluated six months ago may no longer be your best option — and the price changes often as much as the capability. Building a lightweight evaluation habit means you are never more than 48 hours behind the best available tools for your business.