Claude Sonnet 5 vs GPT-5.2-Codex
Claude Sonnet 5 and GPT-5.2-Codex both launched within months of each other (June 2026 and January 2026) and both target serious engineering work, but they're built differently. Sonnet 5 is a general-purpose frontier model with a 1M-token context window and adaptive reasoning effort up to 'max'. Codex is OpenAI's coding-specialized fork of GPT-5.1, tuned for both live pair-programming and long unsupervised execution, with mandatory reasoning and a smaller 400K context. Both are available on OpenKey with one API key and a flat 3% fee over provider list price.
Spec vs spec
| Spec | Claude Sonnet 5 | GPT-5.2-Codex |
|---|---|---|
| Context window | 1M | 400K |
| Max output | 128K | 128K |
| Input modalities | text, image, file | text, image |
| Output modalities | text | text |
| Released | Jun 30, 2026 | Jan 14, 2026 |
| Reasoning | optional | always on |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
anthropic/claude-sonnet-5
Input · 1M tokens
$2.00 + 3%$2.06
Output · 1M tokens
$10.00 + 3%$10.30
Cache read · 1M tokens
$0.200 + 3%$0.206
Cache write · 1M tokens
$2.50 + 3%$2.58
FEE — FLAT, EVERY MODEL3%
openai/gpt-5.2-codex
Input · 1M tokens
$1.75 + 3%$1.80
Output · 1M tokens
$14.00 + 3%$14.42
Cache read · 1M tokens
$0.175 + 3%$0.180
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
anthropic/claude-sonnet-5Cheaper
$41.20
$40.00 provider + 3%
openai/gpt-5.2-codex
$46.87
$45.50 provider + 3%
Pricing on a real workload
Provider list price: Sonnet 5 is $2.00/M input, $10.00/M output. Codex is $1.75/M input, $14.00/M output — cheaper on input, more expensive on output. On OpenKey that's Sonnet 5 at $2.06/M input and $10.30/M output (2.00 x 1.03, 10.00 x 1.03), and Codex at $1.8025/M input and $14.42/M output (1.75 x 1.03, 14.00 x 1.03).
Run the math on a concrete job: 10M input tokens + 2M output tokens. Sonnet 5 costs $40.00 total; Codex costs $45.50. The input price ratio is 1.14x (Codex cheaper per input token), but Codex's output price more than makes up for it once you generate real volume of code. If your workload is output-heavy — long completions, big diffs, generated files — Sonnet 5's output rate makes it the cheaper option overall.
Coding and agent performance
Sonnet 5 posts an Artificial Analysis coding index of 71.5, an intelligence index of 53.4, and an agentic index of 46.7 — no Design Arena scores are published for it in this dataset. Codex has no Artificial Analysis scores here, but it does have Design Arena agent benchmarks across five categories: godotgamedev (elo 1187, rank 12), androidnative (elo 1176, rank 15), mobileapps (elo 1172, rank 24), webapps (elo 1125, rank 22), and fullstack (elo 1060, rank 27). Codex's best category (godotgamedev) beats its worst (fullstack) by 127 elo points, which tells you its agent performance is uneven across task types — check the category that matches your use case before committing.
Context and long-document work
Sonnet 5 supports 1,000,000 tokens of context; Codex supports 400,000 — a 2.5x gap. Both cap max completion output at 128,000 tokens, so the difference is entirely about how much you can feed in, not how much comes back. For tasks like reviewing a full monorepo, ingesting long specs, or multi-file refactors that need the whole codebase in context at once, Sonnet 5's window gives you more room before you have to chunk or summarize.
Modality and reasoning controls
Sonnet 5 accepts text, image, and file input (files being the differentiator) and outputs text only. Codex accepts text and image input, text output — no file modality. Both support tool calling and structured outputs. On reasoning, Sonnet 5 offers five effort levels (low, medium, high, xhigh, max) and reasoning is optional; Codex offers four levels (low, medium, high, xhigh) but reasoning is mandatory — you can't turn it off. If you need a model that always thinks before responding, Codex enforces that by design; if you want the option to skip reasoning for cheap, fast calls, Sonnet 5 gives you that flexibility.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| Long autonomous coding agent runs | GPT-5.2-Codex | Purpose-built for independent execution of complex engineering tasks, with published agent benchmarks across five Design Arena categories |
| Large monorepo or multi-file refactor | Claude Sonnet 5 | 1,000,000-token context vs Codex's 400,000 — 2.5x more room before you need to chunk |
| Output-heavy generation (long completions, big diffs) | Claude Sonnet 5 | $10.30/M output on OpenKey vs Codex's $14.42/M — cheaper once output volume climbs |
| Godot or Android app scaffolding via agent | GPT-5.2-Codex | Ranks 12th (godotgamedev) and 15th (androidnative) on Design Arena, its two strongest categories |
| Processing PDFs or file uploads directly | Claude Sonnet 5 | Only one with file input modality — Codex is limited to text and image |
| Interactive pair-programming with flexible reasoning cost | Claude Sonnet 5 | Reasoning is optional (five effort levels including none), so you can dial cost down for simple edits |
Questions
- Which is cheaper for a typical 10M-input, 2M-output job?
- Claude Sonnet 5, at $40.00 total versus $45.50 for GPT-5.2-Codex. Sonnet 5's output price ($10.30/M on OpenKey) is lower than Codex's ($14.42/M), which outweighs Codex's cheaper input rate once output volume climbs.
- Does GPT-5.2-Codex have Artificial Analysis benchmark scores?
- Not in this dataset — only Design Arena agent scores are available for Codex, covering five categories with elo ranging from 1060 (fullstack) to 1187 (godotgamedev). Sonnet 5 has Artificial Analysis scores (coding index 71.5) but no Design Arena entries here.
- How much bigger is Sonnet 5's context window?
- 2.5x larger: 1,000,000 tokens versus Codex's 400,000. Both models cap output at 128,000 tokens, so the gap only affects how much you can feed in per call, not how much text comes back.
- Can I turn off reasoning on GPT-5.2-Codex to save cost?
- No — reasoning is mandatory on Codex, with a minimum effort level of low and up to xhigh. Sonnet 5's reasoning is optional, with five effort levels (low through max), giving you more control over cost on simple requests.