Claude Opus 5.5 vs Codex: which coding agent to pick (and what it costs)

Checked: October 25, 2026. All prices are US list prices, excluding tax. Subscription numbers come from each vendor's pricing page. API token prices come from Anthropic's Opus 5.5 overview and OpenAI's GPT-6 model pages. Benchmark numbers that aren't from a vendor are labelled as third-party. We did not run private tests. This page has no affiliate links. For every plan in one table, see our AI subscription prices page; for a budget-first view, see which AI subscription to pick.

TL;DR

| Question | Short answer | |---|---| | Same ~$20/mo plan? | Yes. Claude Pro ($20 monthly / $17 annual) includes Claude Code + Opus. ChatGPT Plus ($20) includes Codex on web, CLI, IDE and iOS. | | Cheapest path to try? | Codex: Free / Go ($8) with GPT-6 Luna (limited). Claude Code starts at Pro — Free does not include it. | | API list price (flagship) | Opus 5.5: $4 / $20 per MTok in/out. GPT-6 Astra: $10 / $50. GPT-6.1 Sol (Codex Plus default coding model): $2 / $10. | | Published Terminal-Bench | Third-party Vals AI (neutral agent): Opus 5.5 61.62% / $19.07 per task; GPT-6 Astra 57.07% / $8.21. Official product leaderboard: Codex+Astra 58.2%, Claude Code+Fable 5.1 57.9% (within error). | | Pick Claude Code when… | You want Opus as the daily driver, terminal/CI-first workflows, or a Google Cloud / Bedrock / Foundry route. | | Pick Codex when… | You want a free/cheap entry tier, parallel cloud tasks from GitHub/Slack, an open-source CLI, or lower cost-per-task on published independent runs. |

---

What each product is

Claude Code is Anthropic's agentic coding tool. It reads a repo, edits files and runs commands across terminal, IDEs, desktop, web and mobile. On paid Claude plans it shares one usage pool with chat. Claude Opus 5.5 (claude-opus-5-5) is the current Opus model: 1M context, 128K max output, adaptive thinking always on, released September 22, 2026.

Codex is OpenAI's coding agent inside ChatGPT. It runs on web, CLI, IDE extension, iOS and Codex cloud. ChatGPT Work and Codex share the same plan limits and credits. Current Codex models include GPT-6 Astra (most capable), GPT-6.1 Sol / GPT-6 Sol (coding balance) and GPT-6 Luna (cheap/high volume). GPT-5.5 retires from Codex with ChatGPT sign-in on October 14, 2026 at noon PDT; API-key Codex is unaffected.

---

Subscription access (individual plans)

| | Claude (Claude Code) | ChatGPT (Codex) | |---|---|---| | Free | Chat yes; Claude Code: No | Limited Codex (GPT-6 Luna in desktop app, subject to rollout) | | Entry | — | Go $8/mo (GPT-6 Luna) | | Individual | Pro $20/mo ($17/mo billed annually) — includes Claude Code + Opus | Plus $20/mo — Codex on web/CLI/IDE/iOS; GPT-6.1 Sol, Sol, Luna | | Power | Max from $100/mo (5x or 20x Pro usage) | Pro from $100/mo ($100 / $200 / $500 tiers) | | Team seat | Team Standard $25 monthly / $20 annual | Business $25 monthly / $20 annual (2+ users) |

Sources: claude.com/pricing, developers.openai.com/codex/pricing. Matches our subscription-prices.json hand-check.

Limits matter more than the sticker. Claude usage resets on a rolling 5-hour window, with weekly caps on paid plans; chat and Claude Code share one pool. Codex Plus estimates 15–160 local GPT-6.1 Sol messages per 5 hours (and 5–45 for Astra); Pro currently has no 5-hour limit. Cloud tasks and higher speed modes burn allowance faster. Both let you keep going with usage credits / API keys after the cap.

---

API token prices (pay-as-you-go)

Use these when you run the agent on an API key, or to estimate how expensive a long agent run is.

| Model | Input / MTok | Cached input / MTok | Output / MTok | Source | |---|---|---|---|---| | Claude Opus 5.5 | $4 | $0.20 | $20 | Anthropic Opus 5.5 overview | | Claude Sonnet 5.5 | $2 | $0.20 | $10 | Anthropic pricing | | GPT-6 Astra | $10 | $1 | $50 | OpenAI GPT-6 Astra docs | | GPT-6.1 Sol | $2 | $0.10 | $10 | OpenAI GPT-6.1 Sol / pricing | | GPT-6 Sol | $2 | $0.20 | $10 | OpenAI GPT-6 Sol docs | | GPT-6 Luna | $0.10 | $0.01 | $0.50 | OpenAI models overview |

Opus 5.5 also lists 5m/1h cache writes at $5 / $8 per MTok and Batch at 50% off. OpenAI Batch/Flex are 50% of Standard; Fast is 2x. Prompts over 272K tokens on GPT-6 models use higher long-context rates. For a monthly estimate from your own token volume, use our LLM cost calculator.

Subscription vs API: a $20 plan is usually cheaper if you stay inside the rolling window. Spiky or CI usage often costs less on the API, because you only pay for tokens used. Codex on an API key gets CLI/SDK/IDE but not cloud features (GitHub code review, Slack, etc.).

---

Published benchmarks (not our tests)

We cite three layers. Vendor numbers are upper bounds. Independent runs are closer to like-for-like. Product leaderboards mix agent + model.

Terminal-Bench 4.0

| Model / pair | Score | Cost per task | Who reported it | |---|---|---|---| | Claude Opus 5.5 | 66.4% | — | Anthropic (vendor) | | Claude Opus 5.5 | 61.62% | $19.07 | Vals AI via NeuralTrust summary (third-party, updated ~Sep 27, 2026) | | GPT-6 Astra | 57.9% | — | OpenAI (vendor) | | GPT-6 Astra | 57.07% | $8.21 | Vals AI via NeuralTrust summary (third-party) | | Claude Sonnet 5.5 | 70.6% vendor / 53.03% Vals | $19.33 (Vals) | Gap is large — treat vendor as upper bound | | Codex + GPT-6 Astra | 58.2% (±2.8) | — | Official Terminal-Bench 4.0 leaderboard (product pair) | | Claude Code + Claude Fable 5.1 | 57.9% (±3.8) | — | Same leaderboard (no Claude Code + 5.5 entry listed as of Sep 29, 2026 in the NeuralTrust write-up) |

Vals runs every model in one neutral agent (Terminus 2). NeuralTrust notes Vals flagged Anthropic provider fallback on some Opus rows. Official product scores are within each other's error bars.

Other published tables (vendor-picked)

CursorBench 4.0 (Cursor, Sep 10, 2026 write-up): Opus 5.5 57.8%, Sonnet 5.5 55.5%; no GPT-6 entry yet in that snapshot. DeepSWE / FrontierCode tables exist on vendor blogs but often compare against older Claude or incomplete GPT-6 rows — we don't use them to crown a winner.

---

When to pick which

Choose Claude Code (Opus 5.5) if: - You already live in the Claude ecosystem and want Opus for long agent sessions. - Terminal-first + CI (GitHub Actions / GitLab) matter more than ChatGPT-shaped cloud tasks. - You care about Anthropic's lower flagship list price ($4/$20 vs Astra's $10/$50) and will manage effort / caching so token volume stays reasonable. - You need Claude on Bedrock, Google Cloud or Microsoft Foundry.

Choose Codex if: - You want a free or $8 foot in the door before committing $20. - Parallel Codex cloud tasks from GitHub, GitLab, Linear or Slack are the workflow. - You want the open-source Codex CLI you can audit. - Published independent cost-per-task (Vals) matters: Astra came in under half of Opus per Terminal-Bench task in that run. - Your team is already on ChatGPT Plus / Business.

Many teams run both. Project instructions can live in AGENTS.md (Codex natively; Claude Code can read it) alongside CLAUDE.md. Two agents mean two permission systems and two bills — set approval defaults before you scale either.

---

Caveats

We did not benchmark either agent on your repos. Numbers above are dated published sources. Pilot on a real PR before you standardize.
Quota burns faster than message counts suggest. Effort level, repo size, MCP servers and cloud vs local all move the needle. See the limits section of our budget guide.
Security defaults matter more than brand. Both tools have shipped critical fixes for repository-borne attacks. Prefer least-privilege sandboxes and human approval for risky actions.
Prices and model IDs change. Re-check the linked official pages if this guide is more than a few weeks old.

---

Sources (checked 2026-10-25)

1. Anthropic — Claude Opus 5.5 overview (API pricing $4/$20, specs): https://platform.claude.com/docs/en/models/opus-5-5/overview 2. Anthropic — Claude plans (Pro $20 / $17 annual; Max from $100; Claude Code on paid): https://claude.com/pricing 3. OpenAI — Codex pricing (Free / Go $8 / Plus $20 / Pro from $100; message estimates): https://developers.openai.com/codex/pricing 4. OpenAI — GPT-6 Astra API pricing ($10/$50): https://developers.openai.com/api/docs/models/gpt-6-astra 5. OpenAI — GPT-6 Sol API pricing ($2/$10): https://developers.openai.com/api/docs/models/gpt-6-sol 6. OpenAI — API pricing table (incl. GPT-6.1 Sol $2/$10, Luna $0.10/$0.50): https://developers.openai.com/api/docs/pricing 7. OpenAI — Introducing GPT-6.1 Sol: https://openai.com/index/introducing-gpt-6-1-sol/ 8. NeuralTrust — Codex vs Claude Code (2026) benchmarks (third-party summary of Vals AI, official leaderboard, surveys): https://neuraltrust.ai/blog/codex-vs-claude-code-benchmark 9. StackLoom — AI subscription prices (hand-verified): https://stackloom-ai.pages.dev/ai-subscription-prices/