GPT-5.5 Pro: $30/$180 per 1M Tokens, 88.7% SWE-bench (2026)
GPT-5.5 Pro: OpenAI's top-accuracy mode (2026), 1M-token context, 88.7% SWE-bench Verified, $30 input / $180 output per 1M tokens, parallel test-time compute.
GPT-5.5 Pro is OpenAI's highest-accuracy inference mode for GPT-5.5, released April 24, 2026, running the same 88.7% SWE-bench Verified, 1M-context Mixture-of-Experts model with parallel test-time compute. It costs $30 per 1M input tokens and $180 per 1M output tokens, a 6x premium over standard GPT-5.5's $5/$30 rate, reserved for correctness-critical work.
GPT-5.5 Pro, released by OpenAI on April 24, 2026, is the highest-accuracy inference setting for GPT-5.5, scoring 88.7% on SWE-bench Verified with a 1M-token context window. It costs $30 per 1M input tokens and $180 per 1M output tokens, a 6x premium over standard GPT-5.5, using parallel test-time compute for correctness-critical tasks.
Provider: OpenAI · Family: GPT-5.5
Context window: 1,000,000 tokens · Max output: 128,000
Input modalities: text, image, audio, video, tool-calls · Output: text, tool-calls
About GPT-5.5 Pro
GPT-5.5 Pro is OpenAI's highest-accuracy inference setting within the GPT-5.5 family, released alongside the standard GPT-5.5 model on April 23, 2026, with API access opening April 24, 2026. It is not a separate set of weights: GPT-5.5 Pro runs the same Mixture-of-Experts transformer as GPT-5.5, with 128 expert routing and sparse activation per token, but applies parallel test-time compute so the model explores multiple reasoning paths before answering. It sits at the top of the GPT-5.5 lineup, above the standard GPT-5.5 model and GPT-5.5 Instant, and is positioned for correctness-critical work rather than everyday chat. On published benchmarks, GPT-5.5 Pro shares the same headline scores as GPT-5.5: 88.7% on SWE-bench Verified (443 of 500 real GitHub issues resolved end to end) and 92.4% on MMLU. GPT-5.5 also posts 82.7% on Terminal-Bench 2.0, ahead of Gemini 3.1 Pro's 68.5% and Claude's roughly 65%, and 51.7% on FrontierMath Tiers 1-3 with 35.4% on Tier 4. On Humanity's Last Exam, GPT-5.5 scores 41.4%, behind Claude Opus 4.7's 46.9% and Gemini 3.1 Pro's 44.4%, indicating raw academic-recall reasoning is not where the Pro premium pays off most. Against Gemini 3.1 Pro (80.6% SWE-bench Verified) and Claude Sonnet 4.6 (79.6%), GPT-5.5 Pro's coding lead is the clearest differentiator. OpenAI has not independently published separate GPQA Diamond or AIME 2025 scores for the Pro inference setting versus the standard model. The API context window is 1,000,000 tokens with a maximum output of 128,000 tokens per completion. The Codex product caps GPT-5.5 (including Pro) at 400,000 tokens, so sessions needing the full 1M window must go through the Responses API directly. A context surcharge applies once a session's input exceeds 272,000 tokens: the entire session is billed at 2x input and 1.5x output rates, not just the portion over the threshold, which matters for retrieval-heavy workloads. GPT-5.5 Pro processes text, image, audio, and video inputs in a single unified architecture rather than routing between separate stitched-together models. Vision input preserves up to 10,240,000 pixels or a 6,000-pixel dimension without resizing, which improves chart and document reading plus computer-use accuracy. Audio input supports transcription and translation across dozens of languages, but output remains text-only; there is no native audio-out in the chat completions or Responses API. Tool use, function calling, structured outputs, parallel tool calls, web browsing, and code execution via tools all carry over unchanged from GPT-5.5. GPT-5.5 Pro is priced at $30 per 1M input tokens and $180 per 1M output tokens, a 6x premium over standard GPT-5.5's $5/$30 rate. Cached input pricing is not independently published for the Pro tier; applying OpenAI's standard 90% prompt-caching discount ratio (as published for base GPT-5.5 at $0.50 per 1M cached input against a $5 input rate) would put Pro cached input around $3 per 1M tokens, though this figure is an estimate rather than a confirmed rate. A single 50,000-token correctness-critical code review (40K in, 10K out) costs roughly $3.00. A 1,000-call-per-day workload at the same ratio runs about $3,000/day, underscoring why OpenAI markets Pro for hardest-question accuracy rather than high-volume use. GPT-5.5 Pro is available through the OpenAI API with an API key, and reached general availability on Amazon Bedrock on June 1, 2026 as part of a $50 billion AWS-OpenAI partnership announced April 28, 2026 that ended OpenAI's prior Azure exclusivity; Bedrock pricing matches OpenAI's first-party rates with no markup. It also remains available through Azure OpenAI Service. Inside ChatGPT, GPT-5.5 Pro is restricted to Pro, Business, and Enterprise plans; Free and Plus users do not see it as a selectable model. The GPT-5.5 system card was updated on April 24, 2026 to cover API deployment safeguards for both GPT-5.5 and the Pro inference setting, with separate evaluations noted where the parallel test-time compute setting could materially change risk posture. OpenAI ran its full pre-deployment safety evaluation suite and Preparedness Framework process, including targeted red-teaming for cybersecurity and biology uplift, and incorporated feedback from roughly 200 early-access partners ahead of release. The model's safety posture is balanced: it refuses clear-harm requests but is not unusually restrictive for legitimate technical or research use. GPT-5.5 Pro is best suited to teams that need the highest achievable accuracy on a specific hard problem and can absorb the cost and latency: high-stakes code review, complex multi-file refactors, scientific or legal document analysis, and one-off research questions where a wrong answer is expensive. It is a poor fit for latency-sensitive chat interfaces, high-volume customer support, or any workload where GPT-5.5 Instant, standard GPT-5.5, Gemini 3.1 Pro, or Claude Sonnet 4.6 would clear the accuracy bar at a fraction of the cost; teams doing academic-reasoning-heavy work (Humanity's Last Exam style tasks) may get better results from Claude Opus 4.7 or Gemini 3.1 Pro at lower cost. GPT-5.5 Pro inherits GPT-5.5's data governance: API inputs are not used for training by default, with an enterprise zero-retention option available. OpenAI states SOC 2 Type II compliance, GDPR compliance, and HIPAA eligibility, with data residency options in the US and EU. Under the EU AI Act, GPT-5.5 (and therefore Pro) is classified as a general-purpose AI model with systemic risk obligations. OpenAI has not independently disclosed a training data cutoff date for GPT-5.5 or GPT-5.5 Pro. GPT-5.5 Pro launched as part of the same release wave as standard GPT-5.5 (April 23, 2026) and GPT-5.5 Instant (May 5, 2026), replacing GPT-5.4 and GPT-5.2 Pro as OpenAI's top-accuracy offering; GPT-5.2 models, including GPT-5.2 Pro, were fully deprecated from ChatGPT by June 12, 2026, with existing conversations auto-migrating to GPT-5.5. Rumors as of late May 2026 point to a GPT-5.6 release later in the year with an even larger context window, though OpenAI has not confirmed a date.
Pricing
$30 per 1M input tokens, $180 per 1M output tokens, a 6x premium over standard GPT-5.5's $5/$30. Cached input is estimated at $3 per 1M tokens by applying GPT-5.5's published 90% prompt-caching discount ratio; OpenAI has not independently confirmed a Pro-specific cached rate. Available only on ChatGPT Pro/Business/Enterprise plans and via the API.
Key Features
- Parallel Test-Time Compute: GPT-5.5 Pro explores multiple reasoning paths before answering, OpenAI's mechanism for squeezing extra accuracy out of the same GPT-5.5 weights on hard questions.
- 1M-Token Context Window: Up to 1,000,000 input tokens and 128,000 output tokens via the Responses API, enough for whole-codebase or large-filing analysis in one call.
- Shared 88.7% SWE-bench Verified Score: Resolves 443 of 500 real GitHub issues end to end, the same headline coding result published for standard GPT-5.5.
- Unified Multimodal Input: Text, image, audio, and video processed in one architecture, with vision preserving up to 10.24MP without resizing for accurate chart and document reading.
- AWS Bedrock General Availability: Reached GA on Amazon Bedrock on June 1, 2026 at OpenAI's first-party pricing, as part of a $50B AWS-OpenAI partnership that ended Azure exclusivity.
Pros
- Shares GPT-5.5's 88.7% SWE-bench Verified score, the strongest published agentic coding result in OpenAI's lineup.
- 1M-token context with 128K max output for whole-codebase or large-document work in a single call.
- Available via OpenAI API, AWS Bedrock (GA June 1, 2026), and Azure OpenAI Service at matching first-party pricing.
Cons
- $30/$180 per 1M tokens is a 6x premium over standard GPT-5.5's $5/$30, with no independently published benchmark showing Pro beats base GPT-5.5 by a specific margin.
- No native audio output despite native audio input, and no published tokens-per-second figure for the Pro setting.
- Locked out of ChatGPT Free and Plus plans; only Pro, Business, and Enterprise tiers can select it in the chat UI.
Benchmarks
- mmlu: 92.4
- swe bench verified: 88.7
Frequently Asked Questions
What is GPT-5.5 Pro and who built it?
GPT-5.5 Pro is OpenAI's highest-accuracy inference setting for its GPT-5.5 flagship model, released April 23, 2026 alongside standard GPT-5.5, with API access opening April 24, 2026. It is not a separately trained model: GPT-5.5 Pro runs the same Mixture-of-Experts transformer as GPT-5.5, which uses 128 expert routing groups with sparse activation per token, but applies parallel test-time compute so it explores multiple reasoning paths before producing an answer. It shares GPT-5.5's headline benchmark scores, including 88.7% on SWE-bench Verified (443 of 500 real GitHub issues resolved end to end) and 92.4% on MMLU. GPT-5.5 Pro sits at the top of OpenAI's GPT-5.5 lineup, above standard GPT-5.5 and GPT-5.5 Instant, and was designed for correctness-critical work like high-stakes code review and complex document analysis rather than everyday chat. It directly targets parity with or superiority over Gemini 3.1 Pro (80.6% SWE-bench Verified) and Claude Sonnet 4.6 (79.6%) on agentic coding. GPT-5.5 Pro is priced at $30 per 1M input tokens and $180 per 1M output tokens, with a 1M-token context window.
How much does GPT-5.5 Pro cost per 1M tokens?
GPT-5.5 Pro costs $30 per 1M input tokens and $180 per 1M output tokens, a 6x premium over standard GPT-5.5's $5 input / $30 output rate. A cached-input price specific to Pro has not been independently published; applying GPT-5.5's published 90% prompt-caching discount (base GPT-5.5 charges $0.50 per 1M cached input against its $5 input rate) would put Pro cached input around $3 per 1M tokens, though this is an estimate. As a worked example, a 50,000-token correctness-critical code review (40K input, 10K output) costs roughly $3.00, while a single complex multi-file refactor at 150K input and 30K output costs about $9.90. By comparison, the same task on standard GPT-5.5 would cost roughly one-sixth as much. There is no published batch-API discount specifically for GPT-5.5 Pro. GPT-5.5 Pro is not available to self-host, so there is no infrastructure-cost alternative; it is API-only at the rates above, available directly from OpenAI, on AWS Bedrock at matching first-party pricing, and on Azure OpenAI Service.
What is GPT-5.5 Pro's context window and max output?
GPT-5.5 Pro has a 1,000,000-token context window via the OpenAI API, with a maximum output of 128,000 tokens per completion. The Codex product caps GPT-5.5 (including the Pro setting) at 400,000 tokens of context, so sessions that need the full 1M window must use the Responses API directly rather than the Codex harness. A context surcharge applies once a session's input exceeds 272,000 tokens: the entire session is then billed at 2x the input rate and 1.5x the output rate, not just the tokens over the threshold, which can significantly increase costs for retrieval-heavy workloads that pass large document sets. Compared to Gemini 3.1 Pro and Claude Sonnet 4.6, both of which typically offer context windows in the 200K-1M range depending on tier, GPT-5.5 Pro's 1M window is at the high end of the frontier-model field as of mid-2026. For document handling, GPT-5.5 Pro accepts large multi-file inputs and PDFs as part of its unified text/image/audio/video input pipeline, but teams working with very large corpora should chunk inputs to stay under the 272K surcharge threshold.
How does GPT-5.5 Pro compare on benchmarks vs Claude Opus 4.7 and Gemini 3.1 Pro?
GPT-5.5 Pro shares GPT-5.5's 88.7% SWE-bench Verified score, ahead of Gemini 3.1 Pro's 80.6% and Claude Sonnet 4.6's 79.6%, making agentic coding GPT-5.5's clearest published advantage. On Terminal-Bench 2.0, GPT-5.5 scores 82.7% versus Gemini 3.1 Pro's 68.5% and Claude's roughly 65%, another clear win. On Humanity's Last Exam, however, GPT-5.5 scores 41.4%, behind Claude Opus 4.7's 46.9% and Gemini 3.1 Pro's 44.4%, so for academic-reasoning-heavy tasks the ranking flips. On FrontierMath, GPT-5.5 posts 51.7% on Tiers 1-3 and 35.4% on Tier 4. Critically, OpenAI has not independently published separate GPQA Diamond or AIME 2025 scores showing GPT-5.5 Pro outperforms standard GPT-5.5, so the specific accuracy gain from the Pro parallel test-time setting over the base model remains unverified on those axes. In practice, a 9-point SWE-bench Verified gap (88.7% vs 79.6%) translates to meaningfully more real GitHub issues resolved without human intervention in an agentic coding pipeline.
Is GPT-5.5 Pro open source or proprietary?
GPT-5.5 Pro is fully proprietary and API-only; OpenAI has not released its weights, and there is no open-weights or open-source variant of GPT-5.5 or GPT-5.5 Pro. It is licensed under OpenAI's standard Terms of Use. Access is available through three channels: the OpenAI API directly (api.openai.com) with an API key, AWS Bedrock, which reached general availability for GPT-5.5 and GPT-5.5 Pro on June 1, 2026 at matching first-party pricing as part of a $50 billion AWS-OpenAI partnership, and Azure OpenAI Service. Inside ChatGPT, GPT-5.5 Pro is restricted to Pro, Business, and Enterprise subscription tiers; Free and Plus users cannot select it in the model picker even though their API keys (if they have one) can still call gpt-5.5-pro directly. There are no quantized or self-hostable variants, and no VRAM requirements apply since the model cannot be deployed on-premises.
What modalities does GPT-5.5 Pro support?
GPT-5.5 Pro accepts text, image, audio, and video as input, all processed within a single unified architecture rather than routed between separate stitched-together models as in earlier OpenAI generations. Vision input preserves up to 10,240,000 pixels or a 6,000-pixel dimension without resizing, improving accuracy on charts, infographics, and documents. Audio input supports transcription and translation across dozens of languages in a single workflow. Output, however, is text and tool-calls only; GPT-5.5 Pro does not produce native audio output, so voice applications need a separate text-to-speech step or OpenAI's Realtime API. Tool use is fully supported: function calling, parallel tool calls, structured JSON outputs, web browsing via tools, and code execution via the Code Interpreter-style tool all carry over unchanged from standard GPT-5.5. Compared to Gemini 3.1 Pro, which also supports broad multimodal input, GPT-5.5 Pro's gap is specifically on the output side, where it lacks native audio or image generation.
Does GPT-5.5 Pro train on user data?
No. GPT-5.5 Pro follows the same data policy as standard GPT-5.5: API inputs and outputs are not used to train OpenAI's models by default. Enterprise customers can additionally enable a zero-retention option for further assurance. OpenAI states SOC 2 Type II compliance and HIPAA eligibility, and the model is GDPR compliant with data residency options in both the US and EU regions. Under the EU AI Act, GPT-5.5 (and therefore the Pro setting) is classified as a general-purpose AI model with systemic risk obligations, which brings additional transparency and risk-management requirements for OpenAI as the provider. ISO 27001 certification status was not independently confirmed at the time of writing. On AWS Bedrock, data handling follows AWS's standard model-isolation practices in addition to OpenAI's policy, meaning prompts and completions are not shared across customer accounts or used for retraining. OpenAI has not independently disclosed a training data cutoff date for GPT-5.5 or GPT-5.5 Pro.
Who is GPT-5.5 Pro best for and who should avoid it?
GPT-5.5 Pro is best for teams that need the absolute ceiling of accuracy on a specific hard problem and can absorb the cost: high-stakes code review and complex multi-file refactors (backed by the 88.7% SWE-bench Verified score), whole-codebase or large-document analysis using the 1M-token context window, and one-off correctness-critical research questions where being wrong is expensive. Enterprises already standardized on AWS Bedrock or Azure can adopt it without new vendor contracts since it's available on both at OpenAI's first-party rates. Teams should avoid GPT-5.5 Pro for high-volume customer support or chat, where the $30/$180 per 1M token price (6x standard GPT-5.5) is hard to justify without a confirmed accuracy gain on routine tasks; GPT-5.5 Instant or standard GPT-5.5 are the better fit there. Latency-sensitive or streaming interfaces should also avoid Pro, since no Pro-specific tokens-per-second figure has been published and parallel test-time compute is inherently slower. Finally, teams whose work resembles Humanity's Last Exam (deep academic reasoning) may get better results from Claude Opus 4.7 (46.9%) or Gemini 3.1 Pro (44.4%), both of which outscore GPT-5.5's 41.4% on that benchmark.