Ultra mode: 4 AIs split the work, then merge results

GPT-5.6 lets you dial how much time and compute it's willing to spend thinking, from low, medium, high, xhigh, all the way to max. Above max sits a new tier: ultra. It's not "one model thinking longer" — it's 4 agents brought in by default, splitting the task up to work in parallel, then merging their results.

Toggle the agent count below to watch this "score vs. time" curve shift toward the top-left (further left is faster, higher is more accurate):

For developers, the API side ships a matching multi-agent beta: a single request lets GPT-5.6 run multiple parallel sub-agents and merge their results, effectively letting you build your own ultra-style experience.

"GPT-5.6 is one of the strongest models we've tested on CursorBench, with solid results in early evaluation. It's a step forward for developers on steerability, intelligence, and overall efficiency."Oskar Schulz, President, Cursor

How coding got stronger: a new high score, and cheaper too

Sol is OpenAI's strongest coding model to date. On the Artificial Analysis Coding Agent Index, it scores 80 at the max tier — a new high for the benchmark, 2.8 points ahead of Claude Fable 5 — while using less than half the output tokens, less than half the time, and roughly a third less cost. The whole family benefits: Terra edges out Fable 5, and Luna beats Opus 4.8, each using roughly a third of the time, about half the output tokens, and roughly a quarter of the cost.

Where the savings come from: the model writes its own small program to handle tool calls

Coding and tool-heavy work used to work like this: every time the model called a tool, the tool's entire output got stuffed back into the conversation for the model to re-read before deciding the next step. More round trips meant more tokens burned. GPT-5.6 introduces Programmatic Tool Calling: the model writes a short program that coordinates multiple tools in memory, discards useless intermediate data, and only brings results back to itself at key checkpoints.

A side benefit: because intermediate data is handled in memory and never persisted, Programmatic Tool Calling naturally satisfies Zero Data Retention (ZDR — nothing is retained once processing is done), making it well-suited for data-sensitive industries like healthcare and finance connecting tool-heavy workflows.

It can look at its own mockups, spot problems, and fix them

GPT-5.6 brings one notable change to interface design: give it a high-level direction, and it can produce interfaces that look good, feel right, and work — but more importantly, thanks to stronger computer-use ability, it can look at what it just rendered instead of just spitting out code, catching visual and functional issues and fixing them before handing the work back.

OpenAI's first example: from a single high-level instruction, GPT-5.6 built Saltwind, a playable sailboat-racing game — a fully playable 3D web game. Below is the game itself, embedded live — click "START REGATTA" to sail it yourself (A/D to steer, W/S to trim sail, spacebar to boost):

The other thread is turning a single sentence into an interactive explainer page. In ChatGPT Work, OpenAI gave three examples: an interactive spirograph, wave interference, and a GPT tokenizer visualization — all web pages that recompute live as you drag a slider. Take the spirograph first: the user just said "make an interactive spirograph to explain how it works," and after 1 minute 12 seconds of thinking, the model produced the page below directly (recorded run):

x = (R − r)·cos(t) + d·cos(((R − r)/r)·t)
y = (R − r)·sin(t) − d·sin(((R − r)/r)·t)

Copying a PPT template — whose copy is more faithful

GPT-5.6 can infer a full design spec from a reference document — layout, fonts, spacing, color, recurring content patterns, including rules buried in the Slide Master — and then apply those conventions consistently to new content. OpenAI shared a comparison of "update the numbers to match the reference": on the same reference template, GPT-5.5 missed key elements in the master slide, while GPT-5.6 followed the reference structure more completely.

① Reference template: the model was asked to update the numbers while matching its layout and master-slide elements. Source: OpenAI official blog

② GPT-5.5 output: missed key elements in the master slide. Source: OpenAI official blog

③ GPT-5.6 output: followed the reference structure more completely. Source: OpenAI official blog

The same improvement shows up in documents and spreadsheets too: more faithfully following complex reference formats, more precise handling of formulas and financial models, and better use of layout, spacing, hierarchy, and page structure. That has a direct impact on knowledge work that requires repeatedly matching an existing format.

"GPT-5.6 is highly effective at the long, complex processes behind building production-grade applications. As one of the models Lovable now uses, it's helped users complete the same tasks with roughly 25% fewer steps and 35-48% fewer tool calls, while improving project success rates and cutting stuck runs by 15%."Fabian Hedin, Co-founder, Lovable

Cybersecurity capability jumped a tier — how OpenAI is gating it

GPT-5.6 is OpenAI's strongest cybersecurity model to date, and it uses fewer tokens doing it. The jump across three benchmarks is stark: ExploitBench (measuring progress from touching vulnerable code to achieving arbitrary code execution) rose from GPT-5.5's 47.9% to 73.5%; ExploitGym (turning real vulnerabilities into working exploits) climbed from 15.1% to 24.9% within a two-hour cap, reaching 33.7% when relaxed to six hours; SEC-Bench Pro (proof-of-concept generation on complex software) rose from 45.8% to 71.2%.

OpenAI's safety verdict on this capability: GPT-5.6 is stronger than prior models in both biological and cyber risk, but neither has crossed the "Critical" risk threshold. In cybersecurity specifically, OpenAI's testing suggests it's better at "finding and fixing vulnerabilities" than at reliably mounting end-to-end autonomous attacks against a hardened target, leaving defenders a window to shore up their systems first. Defensive tasks (secure code review, patching, threat modeling, blue-team exercises) remain the priority direction it supports.

The safeguards are layered: protections trained into the model itself, plus real-time checks, ongoing monitoring, and account-level controls. OpenAI also added a "reasoning monitor" that reviews an entire conversation for potential harm, rather than relying solely on flags from a lower-capability classifier to decide whether to block. Compared with before, Sol's cybersecurity safeguards block potentially harmful behavior at roughly 10x the previous strength; for legitimate use that gets wrongly blocked, ChatGPT and Codex offer a one-click option to retry with a lower-capability model. Before launch, OpenAI ran roughly 700,000 A100e-GPU-hours of black-box automated red-teaming.

The full scorecard: where it beats Claude, where it doesn't yet

The charts OpenAI chose to highlight are, naturally, the ones where it's ahead. Laying out the full benchmark table from the appendix gives a more complete picture: GPT-5.6 Sol has categories where it leads by a wide margin, and categories where Claude Mythos 5 pulls ahead. Below are a few key comparisons by category.

Not just a new model: Codex merges into ChatGPT, plus a new ChatGPT Work that does the work itself

The same day as GPT-5.6, OpenAI also reshaped its product lineup: it launched ChatGPT Work and merged the standalone Codex app into ChatGPT, with GPT-5.6 as the engine driving it. This half of the announcement is just as significant as the model itself.

What ChatGPT Work is: a ChatGPT agent with Codex built in, able to pull context from your various apps, break a goal down into steps, and hand back a finished spreadsheet, slide deck, document, or web page directly — not advice, a finished product. It can schedule work and run multi-step tasks in the background on its own (positioned against Anthropic's Claude Cowork), can use your local files and apps, and can also use a new built-in browser to reach websites and online files; it connects into various systems via Plugins; and there's a new experimental Sites feature — drag in a folder or a zip and it instantly becomes a running website.

The desktop merger: the previously standalone Codex app has merged into a new ChatGPT desktop app (available on both Mac and Windows), with Chat, Work, and Codex all available in every plan, including the free tier. Users who already have the Codex app just need to update to get the new ChatGPT desktop app; the previous standalone ChatGPT desktop app has been renamed ChatGPT Classic.

All-in-one on desktop, available on every plan (including free); the old standalone version is now ChatGPT Classic.

Availability: ChatGPT Work launches today for Pro, Enterprise, and Edu plans, expanding to Plus and Business in the coming days.

When you can use it, and what it costs

GPT-5.6 is live today across ChatGPT, Codex, and the OpenAI API, rolling out globally in stages and reaching full availability within the next 24 hours. Of the three tiers, Sol is the flagship, Terra is a lower-cost tier with performance close to GPT-5.5, and Luna is the fastest and cheapest.

Breakdown: in ChatGPT, Plus, Pro, Business, and Enterprise users get Sol at medium effort and above, and Pro/Enterprise can also select Sol Pro. In ChatGPT Work and Codex, Free and Go use Terra, while Plus and above can choose between Sol, Terra, and Luna and set their own effort tier; the max tier is available in settings to every user who has access to GPT-5.6 at all. ultra is available in ChatGPT Work to Pro and Enterprise, and in Codex to Plus and above.

Pricing (per million tokens)

Cache pricing has also changed: starting with GPT-5.6, writing to the cache costs 1.25x the uncached input price, while reading from the cache still gets a 90% discount. For applications that repeatedly send the same long context, caching is now more predictable and easier to budget for.

Want to go deeper

First, a few real screenshots. Derrick Choi, an OpenAI Codex engineer, brainstormed entirely by voice with Codex, then had GPT-5.6 Sol generate whole pages single-shot, producing 15 working websites/UIs. Below are 4 of them: