What's the Next GPT Model? OpenAI's New 2026 Roadmap

If you only want the headline: the next model is GPT-5.6, it is not official yet, and it is expected within weeks rather than months. The newest confirmed flagship is GPT-5.5, released on 23 April 2026. Everything after that is a mix of strong signals and outright speculation, and this article keeps the two firmly apart.

Two things are worth holding in mind throughout. First, OpenAI has shifted from rare, big-number launches to a stream of point releases, so "the next GPT" is now a moving target. Second, the company is spending at a scale that only makes sense if it intends to keep shipping at this pace for years.

This article is structured to take you from the firmest ground to the least firm. We start with the confirmed timeline and the measurable cadence, both of which are facts. We then move to GPT-5.6, where there is strong evidence but no announcement, and carefully label what is leaked. We cover the money and the compute, which are confirmed and which explain the pace. We contrast OpenAI's calm month with Anthropic's chaotic one. And we close with a prediction that is clearly ours, not OpenAI's. Wherever a figure is not sourced, we have left it out rather than guess.

To work out what comes next, it helps to see how fast the line has moved. The early jumps were years apart. The recent ones are weeks apart. Every date below is from public release records unless flagged as a leak.

The confirmed lineage

• GPT-3: 28 May 2020. The 175-billion-parameter model that made few-shot prompting mainstream.
• GPT-3.5: late 2022, the engine behind the original ChatGPT launch.
• GPT-4: 14 March 2023. First mass-market multimodal model, accepting text and images.
• GPT-4 Turbo: a faster, cheaper, longer-context refresh that became the default workhorse.
• GPT-4o: 13 May 2024. The "omni" model adding native audio, vision and real-time interaction.
• o1-preview: 12 September 2024, the first public model in the reasoning ("o") line.
• o3-mini: 31 January 2025, a cost-efficient reasoning model.
• GPT-4.5: 27 February 2025, a large research-preview model focused on chat.
• GPT-4.1: 15 April 2025, a major upgrade aimed at software engineering.
• o3 and o4-mini: 16 April 2025, the stronger reasoning models across maths, code and science.
• Codex (agent): 16 May 2025, OpenAI's cloud-based software-engineering agent.
• GPT-5: 7 August 2025. A unified system folding fast and reasoning models into one product.
• GPT-5.1: 12 November 2025, shipped as Instant and Thinking variants.
• GPT-5.2: 11 December 2025, aimed at professional knowledge work and long-running agents.
• GPT-5.2-Codex: 18 December 2025, an advanced agentic coding model.
• GPT-5.3-Codex: 5 February 2026, the coding-focused checkpoint, with a fast "Spark" variant on 12 February 2026.
• GPT-5.4: 5 March 2026, a frontier model for reasoning, coding and agents.
• GPT-5.5: 23 April 2026. The current flagship, with GPT-5.5 Pro for heavier work.

Two patterns jump out. The "o" reasoning line and the main GPT line have effectively merged: since GPT-5, reasoning is a mode of the flagship rather than a separate product. And the version numbers now move in small increments, which tells you a lot about what to expect next.

It is worth dwelling on the shape of that curve, because it explains why "the next GPT" feels so different in 2026 than it did even eighteen months earlier. GPT-3 to GPT-4 was almost three years. GPT-4 to GPT-4o was a little over a year. GPT-4o to GPT-5 was fifteen months. Then, from GPT-5 onwards, the gaps collapsed into weeks. The line did not just get steeper; it changed character, from rare landmark launches to a rolling product update schedule.

The reasoning line deserves its own note. The "o" series (o1, o3, o3-mini, o4-mini) was OpenAI's bet that letting a model "think" for longer before answering would unlock maths, science and code. It worked, and the lesson was absorbed. By GPT-5 the company stopped selling reasoning as a separate product and folded it into the flagship as a mode you can dial up or down. That matters for what comes next: the next GPT is unlikely to be a standalone "o5". It is far more likely to be a GPT-5.x or GPT-6 with stronger built-in reasoning baked in.

The Codex line is the third thread. OpenAI launched Codex as a cloud-based software-engineering agent in May 2025, then began shipping coding-specialised checkpoints (GPT-5.2-Codex, GPT-5.3-Codex, the fast "Spark" variant) almost in lockstep with the main flagships. That pairing has become a habit, and it is one of the cleaner predictors that any new GPT-5.6 will arrive with a Codex sibling close behind.

For a closer look at how one of these checkpoints stacked up against the competition, see our GPT-5.3 vs Claude Opus 4.6 comparison.

The cadence is the single best predictor of the next release. Here are the recent gaps:

• GPT-5 to GPT-5.1: about three months.
• GPT-5.1 to GPT-5.2: about four weeks.
• GPT-5.2 to GPT-5.3-Codex: about eight weeks.
• GPT-5.3 to GPT-5.4: about four weeks.
• GPT-5.4 to GPT-5.5: about seven weeks.

The trend is clear. Across the first half of 2026 OpenAI averaged a meaningful new model roughly every four to seven weeks. GPT-5.5 landed on 23 April. On that rhythm, a successor in the second half of June or early July is exactly what you would expect, which is precisely the window the GPT-5.6 leaks point to.

There is a strategic reason for the faster drumbeat. OpenAI has openly leaned towards frequent point releases instead of waiting for one giant leap. It keeps the product feeling fresh, lets the company react to rivals quickly, and spreads safety review across smaller steps. The cost is version-number fatigue, but commercially it has worked: ChatGPT passed 900 million weekly active users in 2026.

The naming itself tells a story. The model many observers expected to be GPT-6 was shipped instead as GPT-5.5, codenamed "Spud" internally. That decision, to hold back the big number and release the work as a point upgrade, is the clearest evidence that OpenAI now treats the GPT-6 label as a marketing event reserved for a genuine generational step, not just the next checkpoint off the training run. Combined with the renaming, it strongly implies more 5.x releases before any 6.

There is also a competitive logic to the cadence. The frontier in 2026 is crowded: Anthropic's Claude line, Google's Gemini, xAI's Grok, and a wave of open-weights models from China. When a rival ships, OpenAI wants to be able to answer within weeks, not quarters. A monthly release rhythm is as much a defensive posture as a roadmap. It means that whatever the next GPT is, it will almost certainly be timed partly around what the competition does, not purely around when the model is "ready".

For the user, the practical implication is that the version number on the box matters less than it used to. The gap between GPT-5.5 and a future GPT-5.6 is likely to be incremental and task-specific rather than transformative. If you are choosing a model today, the cadence means you should expect a better option within a month or two, and you should design workflows that can swap models without a rebuild.

What is actually confirmed

• As of 23 June 2026, GPT-5.6 has no official announcement, no system card, no API model page and no published benchmarks.
• A gpt-5.6 string briefly appeared in OpenAI's internal Codex routing and rollout logs in May 2026, then disappeared.
• Reporting says OpenAI chief scientist Jakub Pachocki described the work as a meaningful improvement over GPT-5.5.

What is leaked or rumoured (treat with caution)

• An internal codename trail, iris-alpha, ember-alpha and beacon-alpha, said to be progressing towards a release candidate.
• A 1.5 million token context window, up from the 1 million GPT-5.5 shipped with. This is behavioural observation by testers, not an official spec.
• A focus on long-horizon agentic and Codex computer-use tasks: better multi-hour coding, less token waste, faster responses.
• Cleaner frontend UI generation from minimal prompts.
• A release window of roughly 15 June to 5 July 2026, inferred from prediction markets and leak timing, not an OpenAI commitment.

The strongest single signal is a market one: a Polymarket contract priced a GPT-5.6 launch before 30 June 2026 at roughly 80 to 89 percent. That is a crowd forecast, not a fact, but combined with the Codex log sighting and the cadence it is a coherent picture.

It is worth being explicit about how leaks like this should be read, because the AI rumour cycle is noisy. A model identifier appearing in routing logs is a meaningful signal: it means OpenAI's own infrastructure has a slot for that name, which is hard to fake and usually precedes a real launch. A reported quote from the chief scientist is weaker, because it is second-hand. A context-window figure derived from testers poking at a Codex environment is weaker still, because behaviour can be misread and limits change between a release candidate and general availability. And a Polymarket price is not evidence at all in the usual sense; it is a crowd's probability estimate, useful for timing but not for features.

Stacking those together, the responsible conclusion is narrow. We can be fairly confident a model called GPT-5.6 exists internally and is close. We cannot confirm its context window, its benchmarks, its pricing, or even that it will keep the name 5.6 when it ships. OpenAI has renamed models late in the cycle before, "Spud" becoming GPT-5.5 being the obvious example, so even the version number is not locked.

On the likely shape of the upgrade, the leaks are at least internally consistent. They all point in the direction of agentic, long-running work rather than raw chat intelligence: longer context for holding a whole codebase in mind, fewer wasted tokens so multi-hour Codex runs are cheaper, faster responses, and cleaner UI generation. That is a coherent product story, and it fits OpenAI's stated ambition to turn ChatGPT into something closer to an operating layer that can drive software end to end.

On pricing, the only honest answer is "probably similar to GPT-5.5, but unconfirmed". GPT-5.5 launched at roughly £4 ($5) per million input tokens and £24 ($30) per million output tokens in the API, with the Pro tier at around £24 ($30) input and £142 ($180) output. If GPT-5.6 follows the recent pattern, expect comparable pricing, possibly with a lower effective cost per task if the rumoured token-efficiency gains are real. But no API page exists yet, so treat any specific figure as inference from the predecessor rather than fact.

A word on benchmarks, since they are usually the first thing people ask for. There are none. No SWE-bench Verified score, no GPQA, no AIME, nothing on any public eval has been released for GPT-5.6, because the model has not been released. Any benchmark table you see circulating for GPT-5.6 right now is either projected from GPT-5.5 or invented. We will not publish numbers we cannot source, and you should be wary of anyone who does.

We track the detail as it firms up in our dedicated GPT-5.6 explainer. The honest summary today: GPT-5.6 is very likely real and very likely imminent, but every number attached to it is provisional.

Naming has become a reliable tell. Expect the next family to mirror GPT-5.5: a standard model plus a heavier deliberative tier and, possibly, a small fast one.

• GPT-5.6 (standard): the most likely next release, on the evidence above.
• GPT-5.6 Pro: a deliberative, higher-compute tier, mirroring GPT-5.5 Pro. Rumoured, not confirmed.
• GPT-5.6 Mini: a cheaper, faster variant. Plausible given OpenAI's pattern, but unconfirmed.
• A new Codex checkpoint: the Codex line has shipped alongside most recent flagships, so a 5.6-era coding model is likely.
• A voice model (rumoured as "GPT-Bidi"): leak chatter points to a bidirectional voice model with interruption handling. This is speculation, cited as a leak only.

The voice angle deserves a little more, because audio has been a recurring OpenAI strength since GPT-4o introduced native speech. Leak chatter ties the next era to a bidirectional voice model, sometimes referred to as "GPT-Bidi", able to handle natural interruptions and overlapping speech the way a human conversation does. If that ships, it would be a meaningful product step rather than a benchmark one, the kind of feature that wins consumers more than it wins leaderboards. But to be clear, this is leak-level speculation with no official confirmation, and we flag it as such.

And the big one: GPT-6. It is real as a plan but undated. As of June 2026 there is no architecture paper, no parameter count, no pricing and no launch date. Notably, the model many people assumed would be GPT-6 shipped instead as GPT-5.5 (codenamed "Spud"), which is the clearest sign that OpenAI prefers steady point releases for now and is holding the GPT-6 label for a genuine generational jump.

Sam Altman has flagged long-term memory as the headline feature of the next generation: a system that recalls your preferences, projects and past conversations across weeks and months, paired with stronger agentic goal decomposition and deeper personalisation. If GPT-6 means anything concrete right now, it means that.

Why does memory loom so large? Because it is the missing piece that turns a chatbot into an assistant. Today's models start each conversation, more or less, from scratch; the "memory" features bolted on so far are shallow compared with what Altman has described. A model that genuinely remembers a six-month project, the decisions you made and why, the files you care about, the tone you prefer, is a qualitatively different product. It is also far harder to build safely, because persistent memory raises privacy, consent and data-retention questions that a stateless chatbot avoids. That difficulty is part of why GPT-6 is undated: the hard problem is not raw intelligence, it is making memory trustworthy.

The second pillar is agency. The recent Codex investments, and the rumoured GPT-5.6 focus on multi-hour autonomous work, are all steps towards models that can be handed a goal and left to pursue it across many tools and many minutes without constant supervision. GPT-6, if and when it arrives, is likely to be defined as much by how reliably it can run unattended as by how well it answers a single question.

There is one more naming wrinkle worth flagging. OpenAI could conceivably skip straight to GPT-6 if a training run delivers a step-change, bypassing further 5.x releases. The cadence argues against it, and the "Spud to 5.5" precedent argues against it, but it is not impossible. Treat any confident claim about the GPT-6 release date with scepticism; as of June 2026 there is genuinely no date, no spec sheet, and no system card.

For the mirror-image picture on Anthropic's side, see our sister piece on what's the next Claude model.

The pace only makes sense against the money behind it. On 31 March 2026 OpenAI closed what it called the largest private funding round in history.

• Valuation: about £671bn ($852bn) post-money.
• Amount raised: roughly £96bn ($122bn) of committed capital, up from a £87bn ($110bn) figure announced earlier.
• Anchor investors: Amazon, Nvidia and SoftBank, with continued participation from long-term partner Microsoft. Reporting put Amazon at about £39bn ($50bn), with Nvidia and SoftBank around £24bn ($30bn) each.

On the revenue side, OpenAI said it was generating around £1.6bn ($2bn) per month, an annualised run rate near £19bn ($24bn). It projects roughly £23bn ($29.4bn) of total revenue for 2026, up from about £10bn ($13bn) in 2025 and £2.9bn ($3.7bn) in 2024. Internal forecasts reported in 2026 stretch much further, with subscription revenue approaching £157bn ($200bn) by 2030 and paying users rising from around 35 million towards 220 million.

The takeaway for "what's next" is simple. A company raising at this scale, with revenue tripling year on year, is not slowing its release cadence. It is buying the runway to accelerate it.

The investor mix is itself a signal. Earlier rounds leaned heavily on Microsoft, which provided both capital and Azure compute. The March 2026 round broadened the base dramatically, bringing in Amazon, Nvidia and SoftBank as anchors. Nvidia's involvement is especially telling: the company that sells the chips is now also funding the buyer, a circularity that has drawn scrutiny but also guarantees OpenAI priority access to the hardware its next models will need. SoftBank's role ties OpenAI's fortunes to the same investor backing the Stargate build-out, aligning the money and the compute behind a single roadmap.

There is a sceptical reading too, and it is worth airing. A valuation north of £670bn ($852bn) on a run rate near £19bn ($24bn) implies enormous expectations baked into the price. Critics point out that OpenAI is spending heavily ahead of revenue, that the compute commitments dwarf current income, and that the path to profitability runs through forecasts (200 million-plus paying users by 2030) that are far from guaranteed. None of that changes the near-term picture for model releases, OpenAI has the cash to keep shipping, but it does mean the pressure to monetise each new GPT is intense. Expect the next models to be pitched hard at enterprises and developers, where the money is, as much as at consumers.

For comparison, this funding context is part of why OpenAI could ship calmly through the same month that a rival had a model pulled. Deep pockets and diversified backers buy resilience as well as compute.

Models need chips and power, and OpenAI's compute commitments are the most concrete signal of all that bigger models are coming. The centrepiece is Stargate, the data-centre build-out run with Oracle and SoftBank.

• OpenAI and Oracle agreed to develop 4.5 gigawatts of additional Stargate capacity in the US.
• Combined with the flagship Abilene, Texas site and CoreWeave projects, planned capacity reaches nearly 7 gigawatts with over £315bn ($400bn) of investment across roughly three years.
• Reporting describes an OpenAI-Oracle cloud commitment worth around £236bn ($300bn), with payments of roughly £47bn ($60bn) a year from 2027.
• A first 1-gigawatt deployment tied to an AMD commitment was slated to begin in the second half of 2026.

Gigawatts of compute do not get committed to keep one model running. They get committed because the roadmap assumes a steady stream of larger, more capable models, the kind of trajectory that eventually justifies a GPT-6.

To put the scale in human terms, a single gigawatt is roughly the output of a large nuclear reactor, enough to power a mid-sized city. OpenAI is contracting for nearly seven of them across its Stargate footprint. That is not the infrastructure of a company planning incremental updates to one model; it is the infrastructure of a company planning to train successively larger frontier models for years, and to serve them to a user base it expects to grow several times over.

The timing matters for the roadmap. The first big tranche of new capacity was slated to begin coming online in the second half of 2026, with the bulk of the contracted build-out landing across 2027 to 2031. That phasing fits neatly with the idea that GPT-5.x point releases run on today's compute, while a genuinely larger GPT-6 waits for the new data centres to switch on. If you want a single reason to bet that GPT-6 is a 2027 story rather than a 2026 one, the compute schedule is it.

There is a risk side to this as well. Committing hundreds of billions to compute years in advance is a bet that demand, and model efficiency, will justify it. If models get dramatically more efficient (using fewer tokens and less compute per task, exactly the kind of gain rumoured for GPT-5.6), some of that capacity could end up serving far more users rather than training far bigger models. Either way, the compute is the clearest material evidence that OpenAI's roadmap extends well beyond the next point release.

Capability is only half the story. The other half is whether a model can stay live, and here OpenAI and Anthropic had very different Junes.

OpenAI has had its share of friction: deprecations, rollbacks of over-eager model behaviour, and the usual debates about safety and sycophancy. But it has not had a frontier model pulled offline by government order. That is the line that matters for "what's next", because a model you cannot ship is worse than a model you have not built.

Anthropic crossed that line. On 12 June 2026, three days after launch, the US government ordered it to suspend Claude Fable 5 and Mythos 5 under an export-control directive citing national security. Anthropic complied by switching both models off worldwide, while publicly disagreeing. The trigger was a narrow cyber jailbreak: getting the model to read a codebase and fix software flaws. The full story is in our report on why Claude Fable 5 was banned.

The detail that ties this back to OpenAI is pointed. Anthropic argued the contested capability was "widely available from other models (including OpenAI's GPT-5.5)". In other words, OpenAI's own flagship was used as the benchmark for what counts as already-normal. OpenAI shipped through the same month untouched.

The episode matters for OpenAI's roadmap even though it happened to a competitor, because it sets a precedent that applies to everyone. A frontier model can now be removed by government order, on national-security grounds, on a timeline measured in hours, over a capability that is contested. Every lab, OpenAI included, now has to plan for that possibility. The next GPT will ship into a world where the regulator is a live variable, not a distant one.

OpenAI's own safety posture has its own history of friction worth noting honestly. The company has rolled back model updates that became too sycophantic, faced criticism over how it handles sensitive content, and weathered debate about the speed of its releases versus the depth of its safety review. The rapid cadence this article describes is itself a point of contention: critics argue that shipping a new model every month leaves less time for thorough red-teaming. OpenAI's counter is that smaller, more frequent steps are easier to evaluate than rare giant leaps. Both can be true.

The key contrast remains the one above. OpenAI has had controversies; it has not had a flagship forced offline by the state. For the next GPT, that distinction is the difference between a product you can rely on and one that might vanish overnight.

The two labs are running different playbooks. OpenAI ships small and often, building a consumer "super app" around ChatGPT and an army of point releases. Anthropic ships larger, less frequent jumps (Opus, Sonnet, and the Fable and Mythos line) with a heavier emphasis on published system cards and safety framing.

June 2026 crystallised the contrast. Anthropic launched its most capable model yet and had it forced offline by a state directive within 72 hours. OpenAI, meanwhile, sat on the cusp of yet another routine point release, GPT-5.6, with its biggest constraints being compute and naming rather than regulators.

Look closer and the philosophies diverge further. OpenAI's roadmap is consumer-led: the explicit ambition is a ChatGPT "super app" that hundreds of millions of people use daily, with the API and Codex as powerful but secondary surfaces. Anthropic's is enterprise-and-safety-led: fewer, deeper models, exhaustive system cards running to hundreds of pages, and a brand built on being the careful lab. Neither is obviously right; they are bets on different futures.

That difference shapes what "next" looks like on each side. For OpenAI, next means another point release tuned for breadth, speed and agentic reach, shipped fast and iterated. For Anthropic, next tends to mean a larger, more deliberate jump accompanied by a heavy safety apparatus, which is part of why a single contested capability could trigger such a dramatic intervention. The Fable 5 recall, in that sense, was the safety-first strategy meeting its limit: the most capable, most carefully documented model in the company's history was also the one a regulator felt able to pull.

For users weighing which ecosystem to build on, the cadence and resilience story cuts towards OpenAI on availability, while Anthropic's depth and documentation appeal to those who want maximum transparency. Many teams hedge by keeping both in their stack, exactly the kind of multi-model approach the Fable 5 episode rewarded.

For a fuller side-by-side of the major models and where each one wins, see our Claude vs ChatGPT, Gemini and Grok comparison, and the parallel view from Anthropic's camp in what's the next Claude model.

Combining the cadence, the Codex log sighting and the prediction-market odds, here is our best read as of 23 June 2026.

• Next up is GPT-5.6, most likely before mid-July 2026, with a GPT-5.6 Pro tier alongside it.
• The focus is agentic and Codex computer-use work, with a larger context window and efficiency gains rather than a dramatic reasoning leap.
• A 5.6-era Codex checkpoint is likely; a voice model and a Mini are plausible but unconfirmed.
• GPT-6 stays undated, gated on long-term memory and the new Stargate compute coming online from late 2026, so a 2027 timeframe looks more realistic than 2026.

How confident are we in each part? The existence and imminence of GPT-5.6 is the strongest call, supported by the log sighting, the cadence and the market odds. The Pro tier is a near-certainty given OpenAI's habit of pairing a deliberative model with each flagship. The agentic focus is well-supported by consistent leaks. The Mini and the voice model are genuine speculation, included because they fit the pattern, not because the evidence is strong. And the GPT-6 timing is the softest of all, an inference from the compute schedule rather than anything OpenAI has said.

What would change our mind? A surprise GPT-6 announcement would upend the cadence read entirely. A long silence past mid-July would suggest GPT-5.6 hit a snag, in testing, in safety review, or in a name change. And any move by a major rival, a big Gemini or Claude release, could pull OpenAI's timing forward as a competitive response. The cadence is the baseline; the competition is the accelerant.

We will update this article the moment GPT-5.6 is confirmed, and revise the prediction if the cadence breaks.

The next GPT model is a new point release, GPT-5.6, not a generational leap. It is unannounced but strongly signalled, and on OpenAI's monthly rhythm it should arrive within weeks. The true next generation, GPT-6, is real but deliberately held back for something bigger, most likely built around long-term memory.

What makes the roadmap credible is the scale behind it: an £671bn ($852bn) valuation, revenue tripling year on year, and gigawatts of new compute under contract. While Anthropic spent June fighting to keep a flagship online, OpenAI's main problem was deciding what to call the next one. That, more than any benchmark, is the story of what comes next.