GPT-5.5 Arrives With Bigger Context, Lower Prices, and a Clear Bet on Agents
OpenAI's new flagship ships today in ChatGPT and Codex with a 1M-token API context and $5/$30 per-million pricing. The real story: it's faster and cheaper per task, not just smarter on paper.
OpenAI launched GPT-5.5 on April 23, 2026, and the headline isn't the intelligence gains — it's the economics. The new model matches its predecessor's per-token latency while using fewer tokens to finish the same work, and ships alongside a 1M-context API tier priced at $5 per million input tokens and $30 per million output tokens (as of April 2026). For teams building agents, that combination changes the math on what's actually deployable.
According to OpenAI's announcement, GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise users across ChatGPT and Codex, with a separate GPT-5.5 Pro tier layered on top for demanding reasoning. The company pitches it as "a new class of intelligence for real work and powering agents, built to understand complex goals, use tools, check its work, and carry more tasks through to completion." Strip the marketing voice and what remains is clear: OpenAI is betting its flagship on agentic workflows, not single-shot prompts.
The technical shape comes from developer swyx's breakdown of the release: 400K context in Codex, 1M in the API, and — more significantly — this is the first OpenAI generation co-designed with Nvidia's GB200 and GB300 NVL72 systems. That hardware-software pairing likely explains the efficiency story. OpenAI notes that Codex itself shaved 20% off its own inference speed during development, a recursive improvement the company will likely highlight as evidence the agentic pitch is working internally.
For builders, the pricing shift matters more than any benchmark uplift. At $5 per million input tokens as of April 2026, long-context agent runs that were prohibitively expensive six months ago become viable production workloads. For buyers evaluating coding copilots, the emphasis on debugging, operating software, and working across tools is where enterprise renewals will hinge. For non-technical readers: this is the first generation where "hand a task to the AI and walk away" is a reasonable default rather than an optimistic stretch.
The 82.7% figure swyx cites as of April 2026 — on what reads as a coding evaluation — is directionally interesting but hard to judge without the full system card. Treat it as a signal, not a verdict, until OpenAI publishes the evaluation details.
The bottom line: If you're building with long-context agents, rerun your cost model against GPT-5.5's April 2026 pricing this week — the breakeven on previously unaffordable workflows just moved.