GPT-5.3-Codex
Language Deprecated Full comparison ↗ openai/gpt-5.3-codex · by OpenAI
· decoder-only-transformer
· Built on GPT-5.2
Pricing — 1 offering(s)
Input tokens
- $1.75 / 1M tokens (input) Current 2026-02-05 → 2027-03-31
Output tokens
- $14.00 / 1M tokens (output) Current 2026-02-05 → 2027-03-31
Cached input tokens
- $0.175 / 1M tokens (cached input) Current 2026-02-05 → 2027-03-31
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how GPT-5.3-Codex fits into a cost-aware routing setup
See how →Capability profile
How GPT-5.3-Codex rates across core capability dimensions, with the task-level evidence behind each rating.
OpenAI's own launch framing: "combining the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2." Reasoning-effort levels low/medium/high/xhigh (developers.openai.com/api/docs/models/gpt-5.3-codex).
SWE-bench Verified (500 human-verified GitHub issues; patch must pass the repo's own tests; no file-location hints): resolves 78% — fails roughly 1 issue in 5. Independent (Vals AI leaderboard, #19 of 72; mini-SWE-agent harness).
Function calling, web search, hosted shell, and skills via the Responses API, same as the gpt-5.2-codex predecessor.
400,000-token context window; max input capped at 272,000 tokens, max output 128,000 — same structural constraint as gpt-5.2-codex.
No dedicated inference-latency figure published on this model's own docs page. Vals AI's SWE-bench Verified leaderboard entry (used for the coding score above, not duplicated here) separately reports a 246.53s median latency for that specific evaluation harness — harness/task latency, not a general inference-speed claim.
Same $1.75/$14 per-1M-token pricing as its predecessor gpt-5.2-codex — no price change despite the capability improvement.
Coding benchmarks
View the full coding benchmark leaderboard →| Benchmark | Score | Harness | Source |
|---|---|---|---|
| swe-bench-verified ↗ | 78.0% 2026-02 | mini-swe-agent | 📊 leaderboard ↗ |
| swe-bench-pro | 56.8% 2026-02 | — | 🏢 vendor ↗ |
Operator guidance
Deprecated: OpenAI announced on 2026-10-01 that gpt-5.3-codex shuts down on 2027-04-01 and recommends gpt-6-sol instead, so start new agentic-coding integrations there. It stays priced until the shutdown. Sourced coding-benchmark scores are on Modelglass's coding leaderboard (78.00% SWE-bench Verified via Vals AI, rank #19/72; 56.8% SWE-bench Pro Public at xhigh effort per OpenAI's launch blog). Use the leaderboard rather than this page for benchmark comparisons.
Use cases
- The most capable agentic coding tasks OpenAI currently offers — its own framing: "the most capable agentic coding model to date"
- Long-horizon, tool-heavy coding workflows in Codex or similar agentic environments
Limitations
- Classified 'High' capability for cybersecurity under OpenAI's Preparedness Framework — the first model OpenAI treats as High, and described as the most cyber-capable model it has deployed to date. Source is the GPT-5.3-Codex System Card (openai.com/index/gpt-5-3-codex-system-card/), NOT the API model docs page, which carries no such text (citation corrected 2026-09-03)
- Only the v1/responses endpoint is supported — Chat Completions, Assistants, Batch, and most other OpenAI endpoints are not available for this model
- No standalone general-purpose "GPT-5.3" model exists — this is a coding-specialized model, not a variant of a general-purpose GPT-5.3. (The gpt-5.3-codex API docs page itself is live as of 2026-09-03 — an earlier note that it 404'd is no longer accurate.)