Skip to content

Latest commit

 

History

History
87 lines (61 loc) · 6.05 KB

File metadata and controls

87 lines (61 loc) · 6.05 KB

GPT-5.5

ZH EN Home

One-line take: GPT-5.5 was OpenAI's frontier agentic model of spring 2026 — designed to carry multi-step tasks through to completion with tool use and self-checking — and was succeeded by GPT-5.6 on July 9 2026 (see Post-Launch Landscape below).

Quick Read

Item Conclusion
Vendor OpenAI
Route Frontier agentic model powering Codex, ChatGPT, and API surfaces (succeeded by GPT-5.6, July 2026)
Open source No
Best for High-autonomy coding, research, multi-step agent workflows, and long-context tasks
Main cost 2x the API price of GPT-5.4, and the model boundary blurs with the Codex product boundary
Official website https://openai.com/index/introducing-gpt-5-5/
API reference https://developers.openai.com

Why This Entry Exists

Most entries in this directory are products or tools. GPT-5.5 is a model. It is included because OpenAI explicitly positions it as "a new class of intelligence for real work and powering agents," and because its release reshapes the capability ceiling across multiple existing agent surfaces — Codex, ChatGPT, GitHub Copilot, and any API-based agent builder.

When a model release changes what agents built on top of it can do, the model itself becomes part of the selection landscape.

When To Pick It

  • You want the strongest available OpenAI model for agentic coding tasks through Codex or the API.
  • You need a 1-million-token context window for large codebases or long research workflows.
  • You care about fewer tokens per task — GPT-5.5 completes the same Codex tasks with significantly fewer tokens than GPT-5.4.
  • You are building your own agent system on the OpenAI API and want native tool calling, self-checking, and multi-step planning.
  • Your workflow already runs through ChatGPT, Codex, or GitHub Copilot and you want the latest model upgrade.

When Not To Pick It

  • You want an open-source or self-hosted model. GPT-5.5 is fully proprietary.
  • You are cost-sensitive — input is $5/M tokens, output is $30/M tokens (2x GPT-5.4).
  • You need the current frontier — the landscape has moved past spring 2026 on both sides (GPT-5.6, Claude Fable 5, Opus 4.8); see Post-Launch Landscape below.
  • You want a finished agent product, not a model. In that case, look at Codex (the product) instead.
  • You want local execution without cloud dependency.

Capability Shape

Dimension Assessment Notes
Agentic coding Very strong 82.7% on Terminal-Bench 2.0, highest among public benchmarks at launch
Long context Very strong 1M token context window in API
Tool calling Strong Native tool use for code execution, browsing, file ops
Self-checking Strong Verifies its own output before delivering
Multi-step planning Strong Handles multi-part tasks requiring planning, tool use, and verification
Token efficiency Strong Fewer tokens per task compared to GPT-5.4
SWE-Bench Pro Medium 58.6% — trails Claude Opus 4.7 at 64.3%
Open-source / self-host None Fully proprietary, API and product access only

Where It Runs

GPT-5.5 is not a standalone runtime. It is available through:

  • ChatGPT — Plus, Pro, Business, Enterprise tiers
  • Codex — OpenAI's agentic coding product with sandboxed cloud execution
  • API — Responses and Chat Completions (rolling out)
  • GitHub Copilot — available in the model picker for ask, edit, and agent modes

Operating Cost

Complexity is Medium to High. The model itself is straightforward to use through ChatGPT or Codex, but API-based agent builders face a meaningful cost jump: $5/$30 per million tokens (input/output), exactly 2x GPT-5.4. OpenAI argues the per-task cost may still drop because GPT-5.5 uses fewer tokens to finish the same work, but teams should benchmark this claim against their own workloads.

Relationship To Codex

This profile covers GPT-5.5 as a model. The Codex profile covers the product — the cloud execution environment, isolation model, and review workflow. GPT-5.5 powers Codex but is not the same thing. If you are choosing between agent products, start with Codex. If you are choosing which model to wire into your own agent system, this profile is the right one.

Post-Launch Landscape (as of July 2026)

The frontier moved fast after GPT-5.5's April launch. Three releases changed the model-layer selection picture:

Release Date What it changed
Claude Opus 4.8 (Anthropic) May 28 2026 Fixed Opus 4.7's verbosity and tool-calling issues; added Dynamic workflows (hundreds of parallel subagents in Claude Code); $5/$25 per M tokens
Claude Fable 5 (Anthropic) June 9 2026 First Mythos-class model — a tier above Opus; $10/$50 per M tokens, metered credits in Claude subscriptions since July 7
GPT-5.6 (OpenAI) July 9 2026 Succeeds GPT-5.5 across ChatGPT/Codex/API in three tiers — Sol ($5/$30), Terra ($2.5/$15), Luna ($1/$6) — with an Ultra multi-agent setting; leads the Artificial Analysis Coding Agent Index (Sol 80 vs Fable 5 77.2, GPT-5.5 76.4, Opus 4.8 72.5)

Practical read: GPT-5.6 Sol costs the same as GPT-5.5 and outperforms it, so for new work GPT-5.5 is effectively a legacy choice; the real July 2026 model-layer decision is GPT-5.6 tier selection versus Fable 5 credits versus Opus 4.8 as the dependable default.

Bottom Line

GPT-5.5 was a genuine capability jump for OpenAI's agent ecosystem — strongest agentic coding benchmarks at its April launch, 1M context, and better token efficiency. As of July 2026 it has been superseded by GPT-5.6 at the same flagship price; this profile stays in the map as the reference point for how the spring 2026 model race reshaped agent selection.