OpenAI launches GPT-5.6 and ChatGPT Work

The most consequential thing about GPT-5.6 is not what it can do - it is how much it costs to make it do anything. OpenAI led the launch with token efficiency rather than a headline capability jump, pricing Sol at $5 in and $30 out per million tokens, Terra at $2.50 in and $15 out, and Luna at $1 in and $6 out ^[5]. Sam Altman touted Sol as 54% more token efficient on agentic coding tasks ^[5], and OpenAI said the family was explicitly trained to get more useful work from every token ^[2]. This is a deliberate reframing of the frontier from raw intelligence toward performance per dollar.

The reason for the pivot is visible in the market. Enterprises scaling AI have been hit with what Counterpoint Research's Neil Shah called bill shocks from rising token consumption ^[2], and the Ramp AI Index for May 2026 showed Anthropic leading OpenAI in business adoption for the first time, 34.4% to 32.3% ^[8]. When your rival is winning enterprises on value, matching its intelligence at a third of the price is a sharper weapon than beating it on a benchmark. Artificial Analysis pegged Sol at $1.04 per task versus roughly $2.75 for Claude Fable 5 in its Intelligence Index ^[3], and Terra was pitched at about twice as cheap as GPT-5.5 ^[4]. The tiering itself - one model line, three durable capability rungs that advance on their own cadence - lets buyers dial cost against difficulty instead of overpaying a flagship for routine work.

GPT-5.6 shipped alongside ChatGPT Work, an agentic desktop experience that OpenAI designed to complete jobs rather than answer prompts. It operates across applications and files, executes long-running tasks over hours, and produces finished deliverables - spreadsheets, slide decks, documents, websites, and web apps - while connecting to Slack, Microsoft Teams, Google Drive, SharePoint, Notion, email, calendars, CRMs, and project trackers ^[2]. To ship it, OpenAI merged the Codex coding agent into a single unified ChatGPT desktop app for macOS and Windows that houses Chat, ChatGPT Work, and Codex, renaming the older app ChatGPT Classic and adding a Sites beta for building interactive web apps and dashboards ^[7].

The competitive intent is explicit: ChatGPT Work is positioned as OpenAI's direct answer to Anthropic's Claude Cowork, which launched in January 2026 as an agent that plans and executes multi-step tasks autonomously ^[6]. This is the same organizational bet as the May 2026 decision to fold the consumer ChatGPT team and the Codex team into one product group under Greg Brockman ^[1]- the merger of chat and coding under one roof is what made a single agentic app possible. On X, OpenAI framed the launch around an agent that can stay with a project for hours and turn a goal into finished work, and developer sentiment ran strongly positive and momentum-heavy, with the loudest angle being that this is an agentic desktop app rather than another chatbot. The shift it implies for knowledge workers is a move from prompting toward supervising an autonomous agent that has its own cursor and can touch local files, the web, and enterprise systems ^[9].

Strip away the pricing story and the capability picture gets murky, which is exactly the caveat buyers should hold onto. Sol leads the Artificial Analysis Coding Agent Index at 80, 2.8 points above Claude Fable 5 ^[3], but it sits at 59 on the Intelligence Index against Fable 5's 60 ^[3], and it trails Claude badly on SWE-Bench Pro, roughly 64.6% to 80.3% ^[4]. As The Decoder put it, Sol wins the terminal benchmark OpenAI emphasizes while Fable 5 leads the SWE-Bench Pro number Anthropic emphasizes ^[4]- the victory narrative depends entirely on which coding benchmark you weight.

The independent hands-on read is even more pointed. On developer YouTube, the top hands-on review landed on a memorable framing - that Fable 5 behaves like a better manager or co-founder while Sol is a really good worker - and argued Sol feels comparable to Opus 4.8 rather than Fable 5, yet is priced roughly where Opus sits. In stateless one-shot tests Sol appeared to win lopsidedly, but largely because the rival model frequently refused to answer; when both actually responded, quality was near even. That reframes the whole launch: the reviewer even suggested the .6 in the name may hint OpenAI is holding back a larger GPT-6. Practitioner validation over benchmarks was the dominant social theme, echoed by developers reporting large real-world token spend on Sol - a signal that the market is judging this model by output and cost, not leaderboard rank.

For the first time in a flagship OpenAI launch, the constraint on who could use the model came from Washington rather than a waitlist. GPT-5.6 became generally available only weeks after a limited preview, following US government restrictions tied to concerns about its advanced cybersecurity and biology capabilities ^[6]. Those concerns are not abstract: Sol is described as OpenAI's strongest cybersecurity model yet, scoring 73.5% on ExploitBench, up sharply from 47.9% for GPT-5.5 ^[2]. A model that good at finding exploits is a genuine dual-use artifact, and the staggered rollout is the first visible case of frontier capability itself gating commercial availability.

The second-order story is Microsoft. On the same day, OpenAI named GPT-5.6 the preferred model for Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork ^[10], a move that lands amid reports that Microsoft has been leaning on its own MAI models to cut costs and persistent breakup chatter around the partnership ^[5]. Satya Nadella publicly welcomed GPT-5.6 with Work IQ arriving same-day across Copilot Chat, Cowork, M365 apps, GitHub, and Foundry, praising its efficiency without sacrificing quality - a pointed reaffirmation of the alliance at a moment when its durability is being questioned. The launch thus sits at an unusual intersection: a model powerful enough to draw government restraint, wrapped in a partnership reasserting itself in public even as its internal strains leak into the press.