GPT-5.5 Agentic Flagship Model
source post: GPT-5.5 — OpenAI's First Fully Retrained Agentic Flagship
GPT-5.5 — OpenAI's First Fully Retrained Agentic Flagship
Source: instagram · thevibefounder Saved: 2026-04-25 Tags: #llm #openai #coding #benchmarks Display: GPT-5.5 Agentic Flagship Model — OpenAI's first fully retrained base model since GPT-4.5, optimized for agent runtimes, scoring 82.7% on TerminalBench 2.0.
TL;DR
OpenAI shipped GPT-5.5 ("Spud") on April 23 2026 — its first fully retrained base model since GPT-4.5, optimised as an agent runtime rather than a chat model. It hits 82.7% on TerminalBench 2.0 (vs Claude Opus 4.7 at 69.4%) and is live in the API today, but the $5/$30 per-million-token price tag is double GPT-5.4.
What the post showed
Caption: "GPT-5.5 is pushing the frontier forward again. Stronger on coding, multi-step reasoning, tool use, long context workflows."
Key claims from transcript:
- State-of-the-art on 14 benchmarks vs Claude Opus 4.7 on 4.
- TerminalBench score: GPT-5.5 82.7%, Claude Opus 4.7 69.4% — 13-point gap.
- Engineers describe losing access as "a limb amputated."
- A math professor built a full algebraic geometry app in 11 minutes.
- A genomics researcher analysed 28,000 genes in one session.
- The model wrote its own infrastructure upgrade, making itself 20% faster.
- Ships in ChatGPT and Codex.
What it actually is
- What: GPT-5.5 is the first fully retrained OpenAI flagship since GPT-4.5. Every GPT-5.x release between them was a post-training iteration on the same base; 5.5 reworks the architecture, pretraining corpus, and training objectives to target agentic, multi-step workflows rather than single-turn chat.
- Who built it: OpenAI. Internal codename "Spud."
- Status: Stable, GA. Available in ChatGPT (Plus/Pro/Business/Enterprise), Codex, GitHub Copilot, and via API (model IDs
gpt-5.5andgpt-5.5-pro) as of April 24 2026. Supports Chat Completions and Responses APIs. - Why it matters: First OpenAI model positioned primarily as an agent runtime. Stronger context retention across large codebases, better multi-step tool orchestration, 40% fewer tokens than GPT-5.4 (partially offsetting the higher per-token price). GDPval (agentic research tasks) at 84.9%.
- How it compares to alternatives:
- TerminalBench 2.0: GPT-5.5 82.7% > Claude Mythos Preview 82.0% > Claude Opus 4.7 69.4%.
- SWE-bench Pro (real GitHub issue resolution): Claude Opus 4.7 leads at 64.3% vs GPT-5.5 58.6% — Claude still ahead on that specific benchmark.
- Expert-SWE (OpenAI internal, 20-hour coding tasks): GPT-5.5 at 73.1%.
- GPT-5.5 uses 40% fewer tokens than GPT-5.4, making the effective cost closer to the previous generation despite the 2× nominal price hike.
- "Half the cost of frontier coding models" claim in the reel is misleading — it refers to specialised coding-agent products, not GPT-5.4 itself.
Links
- Primary: Introducing GPT-5.5 — OpenAI
- Docs / API reference: GPT-5.5 Model — OpenAI Developers
- Pricing: OpenAI API Pricing
- Related reading:
- TerminalBench scores deep-dive — Interesting Engineering
- GPT-5.5 vs Claude Opus 4.7 benchmark breakdown — MindWired AI
- GPT-5.5 generally available for GitHub Copilot — GitHub Changelog
- CodeRabbit benchmark results
- TechCrunch launch overview
Kickstarter guide
1. Install / sign up — Existing OpenAI API key works. No new credentials needed. Confirm access at platform.openai.com — model should appear under gpt-5.5.
3. One level deeper — Try the Responses API with multi-tool orchestration. GPT-5.5 was specifically trained for this. Wire up a file-read + web-search + code-exec sequence in Codex or via the API and compare clarification rounds vs GPT-5.4 or Claude Haiku 4.5.
4. Gotchas:
- Price: $5 input / $30 output per million tokens — double GPT-5.4 ($2.50/$15). The ~40% token efficiency gain partially compensates but won't cover the diff on short sessions. Use Batch API (50% off, ≤24h turnaround) for non-latency-sensitive routes like bulk document processing.
- Long-context surcharge: Prompts exceeding 272K input tokens are billed at 2× input, 1.5× output for the full session.
- SWE-bench Pro caveat: Claude Opus 4.7 still leads there (64.3% vs 58.6%) — GPT-5.5 doesn't universally win on every coding measure.
- Data residency: Regional endpoints carry a 10% pricing uplift.