GPT-5.6: What We Know vs What’s Leaked (2026)

As of June 19, 2026, GPT-5.6 does not officially exist. There is no model card on OpenAI’s site, no system card, no API listing, no benchmark table, and no pricing. The current shipping flagship is still GPT-5.5, released April 23. Anyone telling you GPT-5.6 is “out” is reading leaks, not release notes.

That said, the rumors are unusually loud and unusually specific. A model identifier surfaced in OpenAI’s Codex routing logs, prediction markets are pricing a near-certain June launch, and OpenAI’s own chief scientist reportedly called the next model “a meaningful improvement.” This piece draws a hard line between the two: what OpenAI has actually confirmed, and what’s still hearsay — plus how the leaked GPT-5.6 would land against Claude Fable 5, Gemini 3.5, and the open-weight wave from China.

What OpenAI has actually confirmed

Strip away the speculation and the official record is short. OpenAI’s public model lineup, as of mid-June 2026, tops out at GPT-5.5 and its Pro tier. Everything below is from primary releases and system cards, not rumor:

• GPT-5.4 shipped March 5, 2026, succeeding GPT-5.2. It cut factual errors versus 5.2 and added native computer use, scoring 75% on OSWorld-Verified against 5.2’s 47.3% (and edging past the ~72% human baseline). Mini and nano variants followed shortly after.
• GPT-5.5.5 shipped April 23, 2026, with API access the next day. OpenAI called it its “smartest and most intuitive” model, built for long-horizon agentic work — planning, tool use, and multi-step execution. Published benchmarks include 82.7% on Terminal-Bench 2.0 and 51.7% on FrontierMath Tier 1–3. It carries roughly a 1M-token context window.

That’s the confirmed baseline. GPT-5.6, if it ships, would be the next sub-60-day increment in a cadence OpenAI has kept all year. The pattern is the strongest evidence for a June drop — but a pattern is not an announcement.

What’s only leaked (treat as unconfirmed)

Here’s where caution matters. Every spec below is vendor-adjacent or community-sourced. None of it is on an OpenAI page.

The Codex log identifier

On May 14, 2026, a researcher known as Haider spotted a single anomalous line in the routing logs of Codex, OpenAI’s coding agent: a mapping entry pointing to a model called gpt-5.6. It was reproducible for only a few minutes before vanishing from later sessions. That’s the one concrete, reproducible data point. It confirms the name exists inside OpenAI’s infrastructure. It confirms nothing about capabilities.

The 1.5M context claim

Multiple developers report probing an unreleased model through Codex via ChatGPT Pro OAuth and observing behavior consistent with a ~1.5 million-token context window — up from the ~1M that GPT-5.5 shipped with. If accurate, that’s about a 50% jump, and it would matter most for coding agents that burn context reading repos, running tests, and comparing diffs. But “behavior consistent with” is a probe result, not a documented limit. Treat 1.5M as a strong rumor, not a number you can quote to your CFO.

Variants and voice

Reporting and leaked screenshots point to a family, not a single model: a standard GPT-5.6, a GPT-5.6 Pro, and likely a Mini, with a Codex-specific build and possibly a chat-latest alias. TestingCatalog reported on June 18 that early Pro builds had surfaced for some subscribers, alongside a “draggable voice bubble” hinting at a voice-mode refresh (separately rumored to involve a new bidirectional audio model). Efficiency leaks suggest a further 10–15% drop in tokens per task on top of GPT-5.5’s gains — plausible given OpenAI’s trajectory, but unverified.

The “meaningful improvement” quote

The most-cited signal is chief scientist Jakub Pachocki reportedly telling staff — in an internal Slack message surfaced by The Information — that GPT-5.6 is a “meaningful improvement” over GPT-5.5. It’s a real journalistic report, not a public OpenAI statement — and “meaningful improvement” is exactly the kind of phrase that means everything and nothing until benchmarks land.

How the rumored GPT-5.6 would stack up

The awkward truth for OpenAI: even if every leak is right, GPT-5.6 lands into a field that didn’t wait. The headline numbers below are published by the competitors — and several are vendor-reported rather than independently audited, which is exactly why a hypothetical GPT-5.6 has a high bar to clear. (For the prior generation’s three-way fight, see our GPT-5 vs Claude 4 vs Gemini 3 comparison.)

*OpenAI publishes Terminal-Bench and FrontierMath figures rather than a SWE-bench Pro number for 5.5. †Anthropic’s SWE-bench Pro scores are first-party numbers run on its own scaffolding; independent evaluators have reported a tighter picture, so treat them as vendor-reported pending neutral replication. **Z.ai and Moonshot released without head-to-head SWE-bench Pro scores; third-party suites are still measuring.

Two things jump out. First, context size is no longer a moat. A leaked 1.5M window would edge the 1M club, but Gemini 3.5 Pro is expected at 2M, so GPT-5.6 wouldn’t own that crown. Second, the agentic-coding benchmark everyone’s chasing is Claude’s — with an asterisk. Fable 5’s 80.3% on SWE-bench Pro is the headline high-water mark, but it’s Anthropic’s own first-party figure, produced on Anthropic’s scaffolding rather than a neutral harness, and independent evaluators have reported a less commanding lead. So if GPT-5.6 can’t clear it, “meaningful improvement” will read as marketing — but the bar itself is still settling. Anthropic also priced aggressively for the launch window — Fable 5 was bundled free on Pro, Max, Team, and Enterprise plans through June 22 — which sets a competitive tone right as GPT-5.6 is rumored to arrive. Our rundown of the new Claude models goes deeper on that lineup.

The China factor

The week of June 12 delivered two open-weight coding models that change the cost math. GLM-5.2 from Z.ai (Zhipu) runs 744B parameters with ~40B active and a 1M-token context, shipped under an MIT license; Kimi K2.7 Code from Moonshot is a 1T-parameter model with 32B active and a 256K window. Both ship open weights, both target developers who want frontier-class code generation without GPT or Claude API bills. Neither published a clean SWE-bench Pro number, so their true standing is unsettled — but the pricing pressure is immediate and real. If you’re weighing the gap between closed US flagships and open Chinese models, our DeepSeek vs ChatGPT breakdown maps the broader dynamic.

What to do right now

If you build on OpenAI: keep production on GPT-5.5, which is documented and stable. Don’t architect around a 1.5M context window that isn’t in any official limit. If you’re evaluating coding assistants this quarter, benchmark against what ships today — Fable 5, Opus 4.8, GPT-5.5, and the open models — rather than waiting on a rumor. Our guide to the best AI coding assistants compares the shipping options head-to-head.

When GPT-5.6 does land, the first things worth checking are the system card (context limit, knowledge cutoff, safety policy), the real SWE-bench Pro / Terminal-Bench numbers, and the API price per million tokens. Those three settle the hype in an afternoon.

الأسئلة الشائعة

Is GPT-5.6 released yet?

No. As of June 19, 2026, OpenAI has not officially released or announced GPT-5.6. There is no model card, system card, API listing, benchmark, or pricing page. The current shipping flagship is GPT-5.5, released April 23, 2026.

When is GPT-5.6 expected to come out?

Reporting and prediction markets point to late June 2026. Polymarket priced roughly an 83% chance of a release in the June 22–28 window (on nearly $1M in volume as of June 15), and broader “by June 30” framings across Polymarket and Manifold run about 80–89%. OpenAI has not confirmed any date, so this is a market expectation, not a company commitment.

Does GPT-5.6 really have a 1.5 million-token context window?

That’s a leak, not a confirmed spec. Developers probing an unreleased model through Codex’s OAuth reported behavior consistent with ~1.5M tokens, up from GPT-5.5’s ~1M. Until OpenAI publishes an official limit, treat 1.5M as a strong rumor rather than a documented number.

What did OpenAI’s chief scientist say about GPT-5.6?

Per a report in The Information, chief scientist Jakub Pachocki told staff in an internal message that GPT-5.6 is a “meaningful improvement” over GPT-5.5. It was an internal remark surfaced by journalists, not a public OpenAI statement, and it came with no benchmarks attached.

How does GPT-5.6 compare to Claude Fable 5?

You can’t compare them on real numbers yet, because GPT-5.6 has no published benchmarks. Claude Fable 5, released June 9, reports 80.3% on SWE-bench Pro (an Anthropic first-party figure, run on its own scaffolding and not yet matched by independent harnesses) with a ~1M-token context at $10/$50 per million tokens. That’s the public bar GPT-5.6 would be measured against on agentic coding.

Will there be GPT-5.6 Mini and Pro versions?

Reporting and leaked screenshots suggest a family — a standard model, a Pro, and likely a Mini, possibly with a Codex-specific build. None of this is confirmed by OpenAI, and the exact lineup at launch (and which variants arrive first) remains unverified.

Should I wait for GPT-5.6 before building?

For production work, no — build on documented models like GPT-5.5 today and migrate once GPT-5.6 ships with a real system card and pricing. Designing around unconfirmed specs is the fastest way to ship a regression when the rumors turn out to be wrong.

خلاصة القول

GPT-5.6 is the most-anticipated model that doesn’t officially exist. The name is real — it leaked from OpenAI’s own Codex logs — and a late-June arrival looks probable. But “probable” is doing a lot of work: there are no official benchmarks, no pricing, and no system card, which means every spec you’ve seen, including the 1.5M context window, is a rumor wearing a confident face. The competitive picture is the clearest part of the story: Claude Fable 5 set a high (if self-reported) agentic-coding bar, Gemini 3.5 Pro is chasing a 2M context, and open models from China are hammering on cost. GPT-5.6 will have to be genuinely better, not just newer. We’ll update this article with verified numbers the moment OpenAI ships it.