What Is GPT-5.2? OpenAI's Model, Price and Context Window Explained

GPT-5.2 is OpenAI's point-release upgrade to the GPT-5 model family, launched on December 11, 2025. It kept the same four-tier structure as the models around it — a fast non-reasoning mode, a reasoning mode, a high-compute Pro mode, and a coding-specialized variant — but raised the context window, changed the per-token price, and pushed several benchmark scores higher. OpenAI's own current model documentation now lists GPT-5.2 as a "previous flagship," since the GPT-6 family has since taken over as the recommended default, but it was the model powering most ChatGPT and API traffic for much of 2026, and searches for its pricing, context window and release date have stayed strong. Here is what GPT-5.2 actually is, what it costs, and how it compares to GPT-5 and GPT-6.

What Is GPT-5.2?

GPT-5.2 is OpenAI's successor to GPT-5.1, positioned at launch as, in OpenAI's words, a model built for "professional work and long-running agents" — spreadsheet and financial-model generation, slide decks, and multi-step projects that run for longer than a single chat turn. It arrived about three weeks after Google shipped Gemini 3 Pro, and multiple outlets reported that an internal "Code Red" memo from Sam Altman pushed the release earlier than OpenAI's original late-December target — a timeline summarized in GPT-5.2's Wikipedia entry. OpenAI executives, including chief executive of applications Fidji Simo, pushed back on the idea that the launch was purely reactive, saying the company had been working on the model "for months."

Functionally, GPT-5.2 is not a new model generation — it is a refinement of the GPT-5 line, in the same way GPT-5.1 and the later GPT-5.4 are. The real generational jump came nine months later with GPT-6 Astra, which OpenAI now recommends in place of GPT-5.2 for new projects.

GPT-5.2 Release Date and Rollout

GPT-5.2 shipped on December 11, 2025. OpenAI rolled it out to paid ChatGPT tiers (Plus, Pro, Business and Enterprise) first, with GPT-5.2 Instant, Thinking and Pro all available at launch; free-tier and ChatGPT Go users received access somewhat later. On the API, the model is identified as gpt-5.2, with the dated snapshot gpt-5.2-2025-12-11, and its training knowledge cutoff is listed as August 31, 2025 on OpenAI's GPT-5.2 model page.

GPT-5.2 did not stay OpenAI's newest model for long. A coding-focused refresh, GPT-5.3-Codex, followed on February 5, 2026, and a broader update, GPT-5.4, reached general release on March 5, 2026, before OpenAI moved to the GPT-6 generation later that year.

GPT-5.2 Pricing: Cost Per Token

On OpenAI's API, the base GPT-5.2 model costs $1.75 per million input tokens and $14.00 per million output tokens, with cached input priced at $0.175 per million tokens (a 90% discount on repeated context). That is roughly 40% more expensive than GPT-5 on both ends of the ledger — OpenAI's current pricing page for GPT-5 lists the original model at $1.25 input / $10 output per million tokens. GPT-5.2 Pro, the highest-compute tier, costs substantially more: $21 per million input tokens and $168 per million output tokens. The GPT-5.2-Codex variant carried the same base pricing as standard GPT-5.2 ($1.75 / $14), though OpenAI's model documentation now lists that specific Codex snapshot as retired in favor of newer coding models.

Independent pricing trackers such as llm-stats.com also reported a Batch API discount of roughly 50% off the standard rate for GPT-5.2 ($0.875 input / $7.00 output) for requests that don't need an immediate response — OpenAI's own docs pages we checked did not list batch rates directly, so treat that figure as a secondary data point rather than an official one.

What a GPT-5.2 API Call Actually Costs

Per-token prices are easier to read as a worked example. A request that sends 10,000 tokens of input (roughly 7,500 words of text, code or document context) and receives a 1,000-token reply from base GPT-5.2 costs about $0.0175 for the input plus $0.014 for the output — just under $0.032 total. Run the same request against GPT-5.2 Pro instead, and the cost jumps to about $0.21 for the input and $0.168 for the output, or roughly $0.378 total — more than 11 times as expensive for the same request, reflecting the extra reasoning compute Pro uses before it answers. Feed GPT-5.2 near the top of its context window — say 350,000 input tokens, the kind of volume you'd see summarizing a large codebase or document set — and the input cost alone rises to about $0.61 before a single output token is generated. That's the trade-off a 400,000-token window creates: it lets you stuff far more context into one call than GPT-4-era models could handle, but the per-call bill scales with it, which is one reason OpenAI prices cached input at a 90% discount for workloads that repeatedly reuse the same long context.

Context Window and Output Limits

GPT-5.2 has a 400,000-token context window with a maximum output of 128,000 tokens per response — the same ceiling OpenAI's current documentation lists for the original GPT-5. (Several outlets covering the December 2025 launch described GPT-5.2's context window as an increase over GPT-5's, so the GPT-5 figure may have been raised since then, or differed across GPT-5 sub-variants at the time; we're reporting what OpenAI's model pages show today.) The GPT-5.2-Codex variant lists the same 400K window but caps a single request's input at 272,000 tokens. For a plain-language breakdown of what a context window actually controls and why it matters for long documents or agent tasks, see our explainer on what a context window is.

GPT-5.2 Variants: Instant, Thinking, Pro and Codex

OpenAI shipped GPT-5.2 in several flavors rather than a single model:

• GPT-5.2 Instant — a fast, non-reasoning model tuned for everyday chat and quick answers.
• GPT-5.2 Thinking — a reasoning model that works through problems step by step before answering, with standard and extended thinking modes.
• GPT-5.2 Pro — a higher-compute reasoning tier aimed at the hardest professional and technical tasks, priced well above Thinking.
• GPT-5.2-Codex — a coding-specialized snapshot built for software-engineering workflows, since superseded by GPT-5.3-Codex and later releases.

OpenAI also maintained a gpt-5.2-chat-latest identifier on the API, pointing to the non-reasoning conversational version used inside ChatGPT itself, which OpenAI periodically updates behind the same name.

The split matters for cost as much as capability. Instant and the base Thinking tier share GPT-5.2's standard $1.75 / $14 per-million-token pricing, so switching between them changes response quality and latency without changing the rate card. Pro is a separate, far more expensive tier precisely because it spends more compute per answer — OpenAI gives Pro users both a standard and an "extended" reasoning selector, trading longer wait times for more thorough answers on the hardest tasks. Codex, meanwhile, was priced identically to base GPT-5.2 rather than at a premium, which was part of why developers adopted it quickly for coding agents before OpenAI moved on to newer Codex snapshots.

GPT-5.2 vs GPT-5: What Changed

GPT-5.2 is priced higher than the original GPT-5 on both input and output tokens — about 40% more on each — but OpenAI and independent benchmark trackers reported meaningful capability gains to go with the higher price. Reported results from OpenAI's launch materials and the benchmark tracker LLM Stats include:

• SWE-Bench Verified: 80.0% for GPT-5.2 vs. 76.3% for GPT-5
• GDPval (matching or beating human professionals): 70.9% win/tie rate for GPT-5.2 Thinking vs. 38.8% for GPT-5
• GPQA Diamond: roughly 93.2% for GPT-5.2 Pro vs. about 85% for GPT-5
• Latency: LLM Stats reported GPT-5.2 running about 18% faster overall, with time-to-first-token on complex extraction tasks dropping from roughly 46 seconds to 12 seconds

Those are third-party-reported figures rather than numbers we independently reproduced, so treat exact percentages as directional rather than guaranteed. Box CEO Aaron Levie separately said GPT-5.2 scored about 7 points better than GPT-5.1 on Box's internal knowledge-work benchmark — a smaller, more incremental jump than the GPT-5-to-GPT-5.2 comparison above, which is expected since GPT-5.1 had already improved on GPT-5.

GPT-5.2 vs GPT-6 Astra and GPT-6 Sol

GPT-5.2 is a different generation from GPT-6. OpenAI's GPT-6 Astra (official model page) launched on September 3, 2026 — nearly nine months after GPT-5.2 — at $10 per million input tokens and $50 per million output tokens, with a much larger 1.05-million-token context window. That makes Astra roughly 6x more expensive than GPT-5.2 on input and 3.5x more expensive on output, in exchange for a context window more than 2.5 times the size and across-the-board benchmark gains that OpenAI positioned as its new frontier tier.

OpenAI also shipped a cheaper sibling, GPT-6 Sol, later in September 2026, priced at $2 input / $10 output per million tokens — closer to GPT-5.2's price band on output, slightly higher on input, but with the same 1.05-million-token context window as Astra. For teams comparing options today, GPT-6 Sol is the closer like-for-like successor to GPT-5.2 on cost, while Astra is the premium tier. OpenAI's current model documentation marks plain GPT-5.2 as a "previous flagship" and points new integrations toward the GPT-6 models instead.

Astra's pricing also works differently once a request gets long: OpenAI bills prompts over 272,000 input tokens at double the standard input and cache rates and 1.5 times the output rate for the whole request, not just the portion past that threshold. GPT-5.2 has no equivalent long-context surcharge, since its entire 400,000-token window is billed at one flat rate — worth knowing if you're deciding whether a bigger window is actually cheaper for a given workload, or just bigger.

GPT-5.2 Benchmarks

Beyond the GPT-5 comparison above, OpenAI and third-party trackers reported the following scores for GPT-5.2 at launch:

• GDPval: 70.9% win/tie rate against industry professionals (GPT-5.2 Thinking)
• SWE-Bench Pro: 55.6%
• GPQA Diamond: approximately 93.2% (GPT-5.2 Pro)
• ARC-AGI-2: 54.2% (Pro) and 52.9% (Thinking)
• AIME 2025: 100% (reasoning tiers)
• FrontierMath (Tiers 1-3): 40.3% (Pro)

As with any vendor- and tracker-reported benchmark, these numbers come from OpenAI's own launch disclosures and the LLM Stats tracker rather than independent, reproduced testing, so they're best read as the figures OpenAI and outside trackers published rather than a verified ranking.

How to Access GPT-5.2

At launch, GPT-5.2 reached ChatGPT Plus, Pro, Business and Enterprise users first, with free-tier and Go-tier access following afterward and remaining subject to usage limits that vary by plan. On the API, developers could call the model directly using the gpt-5.2, gpt-5.2-pro and gpt-5.2-chat-latest identifiers, billed at the per-token rates above. By late 2026, OpenAI's documentation lists GPT-5.2 as a previous-generation model rather than a current recommendation, and the GPT-5.2-Codex snapshot specifically as retired — so new projects are generally steered toward GPT-6 Astra or GPT-6 Sol, while existing integrations pinned to the GPT-5.2 snapshot can generally keep calling it until OpenAI announces a deprecation date.

What's Next

GPT-5.2 was a mid-cycle upgrade, not a generational leap — it sits between GPT-5.1 and GPT-5.4 in OpenAI's own numbering, and OpenAI has already moved its recommended default to the GPT-6 family. If you're choosing a model today, the practical takeaway is: GPT-5.2 remains a reasonable, cheaper option for workloads that don't need a million-token context window, GPT-6 Sol is the nearest modern equivalent on price, and GPT-6 Astra is the model to reach for if you need the largest context window and the newest benchmark scores. Readers building agentic workflows on top of any of these models may find it useful to understand the underlying mechanics first, since OpenAI pitched GPT-5.2 specifically around long-running, multi-step agent tasks rather than single chat replies.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage