GPT-5.5

GPT-5.5 (API model ID gpt-5.5) is OpenAI's frontier model released 2026-04-23 as the paid ChatGPT tiers GPT-5.5 Thinking and GPT-5.5 Pro, with API access following 2026-04-24 (OpenAI Developer Community launch thread). A free-tier GPT-5.5 Instant variant followed on 2026-05-05, replacing GPT-5.3 Instant as ChatGPT's default model (TechCrunch). It sits between GPT-5.4 and the GPT-5.6 family (GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna) in OpenAI's release cadence, and is also referred to by the codename "Spud" in public reporting (Wikipedia: GPT-5.5). OpenAI's own launch framing: "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," with emphasis on agentic coding, computer use, knowledge work, and early scientific research (launch thread).

Key specs

Values as of 2026-04-23 (release), re-verified 2026-07-12 against OpenAI's developer docs and cross-checked against models.dev (as of 2026-07-12) — no discrepancies found.

Spec	Value	Source
Context window	1,050,000 tokens (marketed as "1 million")	Model page; models.dev
Max output	128,000 tokens	Model page
Modalities	Text + image input; text output only. No audio or video	Model page
Reasoning effort	none / low / medium (default) / high / xhigh	Model page
Knowledge cutoff	December 1, 2025	Model page
Pricing, standard	$5.00 / MTok input, $0.50 / MTok cached input, $30.00 / MTok output	Pricing; models.dev
Pricing, GPT-5.5-Pro	$30.00 / MTok input, $180.00 / MTok output	Pricing
Pricing, long-context surcharge	Prompts over 272K input tokens bill at 2x input / 1.5x output rate	Model page
Rate limit, Tier 5	15,000 RPM / 40,000,000 TPM / 15B batch-queue tokens	Model page

Cached reads reuse the model's KV-cache state and cost a tenth of standard input — a first-class lever for prompt-caching strategy in long agent loops, before the 272K surcharge tier changes the math (model page, as of 2026-07-12). Supported tools at API launch: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search (model page).

Benchmark results

Benchmark	GPT-5.5	Source	As of
SWE-bench Verified	82.6%	Vals AI leaderboard	2026-07-12
SWE-bench Pro	59.4%	OpenAI GPT-5.6 announcement, coding eval table	2026-07-12
Terminal-Bench 2.1	83.4% (Codex CLI harness)	tbench.ai leaderboard	2026-07-12

On the Vals AI SWE-bench Verified leaderboard, GPT-5.5 (82.6%) trails GPT-5.6 Sol (96.2%), Claude Fable 5 (95.0%), Claude Opus 4.8 (88.6%), Grok 4.5 (86.6%), and GLM-5.2 (82.8%), and leads Claude Sonnet 5 (79.6%), Kimi K2.7 Code (78.2%), and Kimi K2.6 (76.2%) — all read from the same snapshot (Vals AI, as of 2026-07-12).

Terminal-Bench 2.1 is harness-dependent. GPT-5.5's 83.4% (± CI 81.2–85.6) uses OpenAI's own Codex CLI harness; on the same leaderboard with the generic Terminus-2 harness, GPT-5.5 scores only 78.2% — a 5-point swing from harness choice alone on an identical model (tbench.ai, as of 2026-07-12).

Agentic behavior notes

• Token efficiency. OpenAI states GPT-5.5 "uses significantly fewer tokens to complete the same Codex tasks" than GPT-5.4, while matching it on per-token latency in real-world serving — an efficiency claim, not a specific token-count figure OpenAI itself publishes (launch thread).
• Effort dial as a cost/latency lever. GPT-5.5 exposes five reasoning-effort levels (none / low / medium / high / xhigh, default medium, see Key specs) as a first-class per-request parameter — an operator budgeting agent cost and latency should treat reasoning_effort as a per-task setting rather than a fixed global default (model page).
• Hallucination-reduction claim. OpenAI reports GPT-5.5's individual claims are 23% more likely to be factually correct and its responses contain a factual error 3% less often than GPT-5.4's, measured on de-identified ChatGPT conversations that users previously flagged as containing factual errors — a set OpenAI itself says deliberately targets "especially hallucination-prone cases, not a representative slice of all production traffic." OpenAI also notes GPT-5.5 tends to make more factual claims per response than GPT-5.4, which is why the response-level improvement (3%) is smaller than the claim-level one (23%) (OpenAI Deployment Safety Hub, as of 2026-07-12). Treat it as a narrow, self-selected signal, not evidence that generalizes to an arbitrary factual-grounding task — a case worth reading alongside evaluation awareness before citing it in an LLM-as-judge setup.

Limitations

• Restricted cyber-capability variant. According to trade-press reporting, OpenAI created a separate, gated GPT-5.5-Cyber variant "available only to a limited group of approved users" for penetration testing, vulnerability discovery, exploitation analysis, and malware reverse engineering; access requires an application covering credentials and intended use, plus government-adjacent vetting, with a phased rollout — the base GPT-5.5 covered by this page does not include that tooling (FinTech Grid, accessed 2026-07-12; single-source report, not independently confirmed against an OpenAI primary announcement in this pass).
• No audio or video modality. Input is limited to text and images; output is text-only — a constraint for any agent harness that expects native speech or video handling (model page).
• Hallucination claim is not a general factuality guarantee (see Agentic behavior notes above) — an independent engineering-blog critique finds a widely circulated "60% hallucination cut" figure in press coverage "is not traceable to OpenAI's GPT-5.5 documentation" and instead looks like an averaging of three older, unrelated GPT-5-series figures (a ~6x hallucination reduction for GPT-5 thinking vs. o3, a ~45% factual-error reduction for GPT-5 vs. GPT-4o, and a 33% false-claim reduction for GPT-5.4 vs. GPT-5.2) rather than anything OpenAI reported for GPT-5.5 itself (Wire Blog, accessed 2026-07-12).
• Terminal-Bench 2.1 score requires OpenAI's own harness to reach its headline figure; a generic harness on the identical model scores 5 points lower (see Benchmark results above) — a reminder that a model-only score without stating the harness is an incomplete comparison (tbench.ai).

Related

• gpt-5-6-sol, gpt-5-6-terra, gpt-5-6-luna — the GPT-5.6 family that followed GPT-5.5 in OpenAI's release cadence; GPT-5.6 Sol leads GPT-5.5 by a wide margin on SWE-bench Verified (96.2% vs 82.6%, Vals AI, as of 2026-07-12).
• gpt-5-3-codex-spark — sibling point-release from the GPT-5.3 generation; GPT-5.5 Instant succeeded GPT-5.3 Instant as ChatGPT's default free-tier model in May 2026 (TechCrunch).
• Claude Opus 4.8 — closest-generation Anthropic comparator; the two models trade the lead across different benchmarks rather than one dominating (see Benchmark results).
• claude-fable-5, grok-4-5, glm-5-2, kimi-k2-6, kimi-k2-7-code — other frontier models scored on the same SWE-bench Verified snapshot.
• context-window, context-compaction — GPT-5.5's 1.05M-token window and tiered pricing above 272K tokens are direct inputs to both.