Skip to main content
  • Pricing
Comparisons

Claude Sonnet 4.5 vs GPT-5.3-Codex: Documented Specs, Prices and Retirement Dates

By Oakgen TeamPublished 2 min read

Share
Claude Sonnet 4.5 vs GPT-5.3-Codex: Documented Specs, Prices and Retirement Dates

Claude Sonnet 4.5 and GPT-5.3-Codex are both coding-focused models that their makers have now deprecated. As of 2026-11-19, GPT-5.3-Codex has the larger context window and lower list prices, while Claude Sonnet 4.5 retires first, on November 30, 2026. This page compares only what Anthropic and OpenAI document themselves.

Side-by-side specs

ItemClaude Sonnet 4.5GPT-5.3-Codex
MakerAnthropicOpenAI
ReleasedSeptember 29, 2025February 5, 2026
Context window200K tokens400,000 tokens
Max inputNot listed separately272,000 tokens
Max output64K tokens128,000 tokens
Input price per million tokens$3$1.75
Cached input per million tokens$0.30 (cache read)$0.175
Output price per million tokens$15$14
Input / outputText and images, text outText and images, text out
Knowledge cutoffJan 2025 reliable, Jul 2025 training dataAug 31, 2025
Reasoning controlsExtended thinkingEffort levels low, medium, high, xhigh
StatusDeprecated Sep 30, 2026Deprecated (notice Oct 1, 2026)
RetirementNovember 30, 2026April 1, 2027
Recommended replacementClaude Sonnet 5.5gpt-6-sol

Sources: Anthropic's Sonnet 4.5 model page and models overview; OpenAI's GPT-5.3-Codex model page and deprecations page, all read 2026-11-19. Cache pricing means different things at each vendor: Anthropic bills a cache read at 10% of input, and OpenAI lists a cached-input rate. Check each vendor's pricing page for cache-write charges, batch discounts and long-context rules before budgeting.

A worked cost example

Take a job that sends 1,000,000 input tokens and receives 100,000 output tokens at list price with no caching or batch discount.

  • Claude Sonnet 4.5: 1 x $3 + 0.1 x $15 = $3.00 + $1.50 = $4.50
  • GPT-5.3-Codex: 1 x $1.75 + 0.1 x $14 = $1.75 + $1.40 = $3.15

At these list prices GPT-5.3-Codex costs about 30% less for that mix (3.15 / 4.50 = 0.70). The gap changes with your input to output ratio, because the output prices are close ($15 and $14) and the input prices are not ($3 and $1.75). Tokenizers also differ between vendors, so the same text can count as a different number of tokens in each.

What each vendor says the model is for

OpenAI's launch post calls GPT-5.3-Codex its most capable agentic coding model to date. It says the model combines the coding performance of GPT-5.2-Codex with the reasoning and professional knowledge of GPT-5.2, and runs 25% faster for Codex users. At launch it was available with paid ChatGPT plans across the Codex app, CLI, IDE extension and web, with API access described as coming; the API model page now lists pricing and limits.

Anthropic's September 29, 2025 announcement called Claude Sonnet 4.5 "the best coding model in the world" and the strongest model for building complex agents, at the same $3 and $15 price as Claude Sonnet 4. It also introduced context editing and a memory tool in the Claude API for longer-running agents. Developers called it through the ID claude-sonnet-4-5.

Why the benchmark numbers do not line up

Each vendor publishes scores, but they do not measure the same thing.

  • Anthropic reports 77.2% on SWE-bench Verified (average of 10 trials with a 200K thinking budget) and 61.4% on OSWorld.
  • OpenAI reports 56.8% on SWE-Bench Pro (Public), 77.3% on Terminal-Bench 2.0 and 64.7% on OSWorld-Verified, all at xhigh reasoning effort.

OpenAI's own post notes that SWE-bench Verified tests only Python, while SWE-Bench Pro spans four languages and is designed to be more contamination-resistant. The two OSWorld numbers come from different benchmark versions and different setups. All of these are vendor-reported, so a gap between them is not evidence that one model is better.

Choosing between them today

  • Starting a new project: Both vendors point you to newer models. Anthropic recommends Claude Sonnet 5.5, and OpenAI lists gpt-6-sol as the replacement.
  • Maintaining an existing integration: Sonnet 4.5 needs a migration within days of this page's date. GPT-5.3-Codex gives you until April 1, 2027 by OpenAI's schedule.
  • Large repositories or long transcripts: The documented limits favor GPT-5.3-Codex, with 400,000 tokens of context against 200K.

Pin an exact model ID in production, and re-run your own tests, because published benchmarks do not predict how a model behaves on your code.

Take this further with Oakgen Agent

Run the arithmetic in a sandbox, build a cost comparison, and export it as a spreadsheet; it can also search the web for current vendor prices.

Open this task in Oakgen Agent

The prompt is filled in for you. Nothing is sent until you press send.

Open in Oakgen Agent

Frequently asked questions

Which is cheaper per token, Claude Sonnet 4.5 or GPT-5.3-Codex?

As of 2026-11-19, OpenAI's model page lists GPT-5.3-Codex at $1.75 per million input tokens, $0.175 cached input and $14 output. Anthropic's model page lists Claude Sonnet 4.5 at $3 input, $0.30 cache read and $15 output per million tokens. On those list prices GPT-5.3-Codex is lower on every line, before batch discounts and any long-context surcharges that each vendor may apply.

Do both models still work today?

Yes, but both are deprecated. Anthropic's page says Claude Sonnet 4.5 was deprecated on September 30, 2026 and retires on November 30, 2026, with Claude Sonnet 5.5 as the recommended replacement. OpenAI's deprecations page lists gpt-5.3-codex with a notice date of October 1, 2026, a shutdown date of April 1, 2027 and gpt-6-sol as the recommended replacement.

Can I compare their benchmark scores directly?

Not safely. Anthropic reports 77.2% on SWE-bench Verified and 61.4% on OSWorld for Sonnet 4.5. OpenAI reports 56.8% on SWE-Bench Pro (Public), 77.3% on Terminal-Bench 2.0 and 64.7% on OSWorld-Verified for GPT-5.3-Codex. The benchmarks, versions, harnesses and reasoning settings differ, and all numbers are vendor-reported, so this page does not rank the models on them.

Which model has the larger context window?

GPT-5.3-Codex. OpenAI lists a 400,000-token context window with 272,000 maximum input tokens and 128,000 maximum output tokens. Anthropic lists Claude Sonnet 4.5 at 200,000 tokens of context and 64,000 maximum output tokens.

Sources

  1. OpenAI, Introducing GPT-5.3-Codex, February 5, 2026 (read 2026-11-19)
  2. OpenAI API docs, GPT-5.3-Codex model page (read 2026-11-19)
  3. OpenAI API docs, Deprecations (read 2026-11-19)
  4. Anthropic, Introducing Claude Sonnet 4.5, September 29, 2025 (read 2026-11-19)
  5. Anthropic Claude docs, Claude Sonnet 4.5 model page (read 2026-11-19)
  6. Anthropic Claude docs, Models overview (read 2026-11-19)
Share