GPT-5.6 rumors intensify as OpenAI eyes late-June release

OpenAI will be releasing its new flagship language model GPT-5.6 on June 23rd, as per recently leaked documents in developer forums and prediction markets. The release coincides with a brief period during which Anthropic’s Claude Fable 5 model, which was released on June 9 and recalled by the US Department of Commerce on June 12 due to a security flaw, transitions to a subscription-based service.

The timing of the GPT-5.6 release holds strategic significance for the international AI market. According to Polymarket information available up until June 15, the probability of GPT-5.6’s unveiling between June 22 and 28 was estimated at 83%, which indicates that investors were increasingly certain about OpenAI making swift use of their rival’s misstep.

The prediction market has increasingly converged around a late-June launch. Polymarket contracts concerning GPT-5.6 coming out between June 22 and 28 have been steadily rising since early June, reaching 83% on June 15. Prediction markets are not necessarily indicative of insider information but rather a measure of market sentiment and information aggregation around major technology events.

GPT 5.6 Insights gained from leaks

The developers’ community has pieced together the profile of GPT-5.6 using routing traces available from the internal tests process. According to Android Headlines, the chief scientist of OpenAI, Jakub Pachocki, announced the model as a “meaningful improvement” compared to GPT-5.5, which was rolled out on April 23. Routing traces of the latest version have been found in OpenAI’s Codex backend log since mid-May by several codenames such as “iris-alpha,” “ember-alpha,” and “kindle-alpha”. Kindle-alpha is identified as the release candidate.

Two key capabilities of the model are an improved context window of up to 1.5 million tokens and increased agentic coding capability. As reported by Developer Mark Kretschmann on X, “from what I am hearing, GPT-5.6 is super strong and beats anthropic mythos on many agentic coding benchmarks.” Code generation in the front end has also been highlighted. According to reports, Kindle Alpha can create high-quality UI code with the use of simple prompts and not complex instructions, as in previous models.

However, not all comments have been positive. A developer reportedly has conducted tests comparing Kindle-alpha with Kepler, an earlier version of a checkpoint. Using the same prompts and at the same tier level, he found out that Kindle Alpha is worse than Kepler.

Price war

The biggest commercial aspect revealed in the leaks is the price. Sources show that the API of GPT-5.6 will be around one-third cheaper than that of Fable 5. Anthropic priced Fable 5 at $10 per million input tokens and $50 per million output tokens, which is twice the price of its previous version. Currently, GPT-5.5’s price is $5 and $30 per million input and output tokens, respectively.

The cost difference is crucial when it comes to adoption. In contrast to Claude Fable 5, with costs of $10 per million input tokens and $50 per million output tokens, DeepSeek’s V4-Pro model has prices of $0.435 and $0.87, respectively. The costs for GPT-5.5 are placed somewhere between the two competing models, costing $5 and $30, respectively.

If OpenAI releases a version of GPT-5.6 with performance matching that of Fable 5, or better yet, surpassing its performance levels, while keeping the same pricing structure as GPT-5.5 or even beating it, the company will certainly put more pressure on Anthropic’s bottom line, particularly for enterprises deploying long-running AI agents and coding workflows.

Such pricing strategies of the models illustrate the changing economic environment of AI. Prices on Claude rose 100% from Opus 4.8 to Fable 5 by Anthropic. DeepSeek lowered the cost of V4-Pro permanently after a promotional period and positioned itself as one of the cheapest frontier providers.

According to Wharton Professor Ethan Mollick, who used Fable 5 to perform his tests, the model is able to run without human intervention for about nine and a half hours, though the cost implications of doing so make it a production nightmare.

The economics of agentic AI have become an active research topic.

In recent research carried out by MIT and Stanford University, researchers revealed that autonomous coding agents use up to 1,000 times more tokens compared to regular chat and coding workload. In this case, input tokens make up a larger part of the expense. It was discovered that the greater number of tokens used is not always associated with higher results, meaning that pricing efficiency will soon take its place along with benchmark leadership in enterprise deployments.

An intense month for frontier AI

Three of the most noteworthy AI models may all launch this month: Fable 5, Google’s Gemini 3.5 Pro, and GPT-5.6. It underscores how narrow the release window has become. After just less than three months from GPT-5.4, OpenAI has managed to release a model every seven weeks or so.

Apart from competitive dynamics, there is another layer of pressure from the markets. Anthropic announced its S-1 filing with the SEC on June 1, followed by OpenAI announcing its S-1 filing on June 8. Once SpaceXAI finishes its roadshow next month, the combined market cap of the three firms will be higher than $3.6 trillion. The first mover can set expectations in the market.

There have been some opposing opinions on the launch date of GPT-5.6. On June 16, Kai (@hqmank) commented on X, saying that OpenAI faces an unwinnable situation with respect to the launch date. If OpenAI releases a model with lesser capability than Fable 5, then Anthropic owns the frontier narrative. But if OpenAI releases a better model, then it would face similar regulatory scrutiny as Fable 5. “Model capability is no longer the only bottleneck,” he added.