Claude Code vs Codex: cost and model performance

In November 2026, Claude Pro costs $20 per month, or $17 per month billed annually, while Codex is available on ChatGPT Free and the $20 Plus plan. On model performance, our shared-agent benchmark gives Claude Opus 5.5 a 92.2% solve rate at $1.01 per solve, against 86.8% at $0.30 per solve for GPT-6.1 Sol. Those are API list-price benchmark costs, not subscription costs.

Sources: Claude pricing, Codex plans, Codex pricing and Requesty's coding benchmark.

Which models solve more tasks?

Our coding benchmark, updated October 29, 2026, puts 129 private Go backend tasks through the same agent, with bash, read, edit and write tools. A task counts as solved when every hidden test passes and the existing tests still pass. Each model gets an 80-turn budget, and cost per task includes retries.

Opus solved 5.4 percentage points more tasks than Sol. Sonnet solved 0.8 percentage points more than Sol; at $0.59 per solve, it sits between Sol and Opus on cost in this test.

On the benchmark's 18 extreme-tier tasks, Opus solved 77.8%, while Sol and Sonnet each solved 66.7%. That is a small set of tasks, so test your own difficult work before relying on the gap.

The agent setup is the same across all four rows. Claude Code and Codex have their own tooling and session behavior, so our benchmark compares the models rather than running the two products head to head. For more model results, see our coding model benchmark comparison.

For an in-product comparison, cubic's Bug Bench scored Claude Opus 5.5 inside Claude Code at 1.49 and GPT-6.1 Sol inside Codex at 1.22. A lower relative bug rate is better. This is a separate test from our shared-agent benchmark.

What does that cost per 100 solves?

Assume a workload with the same task mix, solve rates and costs as the benchmark. At the published cost per solve, 100 solved tasks cost 100 x $1.01 = $101 with Opus 5.5 and 100 x $0.30 = $30 with GPT-6.1 Sol, a difference of $101 - $30 = $71.

Sol would also leave more attempts unsolved at the benchmark's observed rate. The calculation uses API list-price benchmark costs, including retries within a task; it is not a Claude or ChatGPT subscription bill. Both model runs had a 100% cache hit rate, so the cost difference may look different on a workload with less cache reuse.

What do the plans cost?

Claude Code comes with paid Claude plans and shares usage limits with Claude chat. Codex is included across ChatGPT plans, including Free and Go, with limits that vary by plan. Sources: Claude pricing and OpenAI's Codex plan guide.

Codex Cloud runs tasks on OpenAI-managed computers and can keep working while your own computer is asleep.

Plan prices and limits come from Claude's pricing page, Anthropic's plan guide and OpenAI's Codex pricing page.

The monthly bill can look different from the plan price. Anthropic reports an average of around $13 per developer per active day and $150 to $250 per developer per month across its enterprise deployments. OpenAI's Codex rate card puts average cost at approximately $100 to $200 per developer per month, depending on the model, concurrent instances, automations and fast mode.

Those are vendor averages across different deployments, not a controlled price comparison. Start with the plan you already have, track how often you hit its limits, and then compare the cost of a higher plan with your expected API usage. For team budgets, see our guide to managing Claude Code, Cursor and Codex costs.

What do Claude Code sessions look like?

Our April 2026 cost note found that Claude Code active users routing through Requesty averaged $108 per month. The median was $23 and P95 was $296: in this sample, half of active users spent $23 or less, while the 95th percentile was $296.

These users chose to route through Requesty, so the sample has selection bias. It is a useful budgeting reference for gateway usage, not the price of a Claude subscription.

The session-depth note counted a median of 16 calls per Claude Code session and 209 at P95 across 594 traces. Trace coverage for Claude Code fell to about 7% in April, so those figures describe a subset of sessions.

Our finish-reason note found that 73% of Claude Code calls finished with tool_calls, while 20% finished with stop. A tool_calls finish means the model returned a tool call; it does not tell you whether the user's coding task was solved.

What can interrupt a session?

Local Codex uses an OS-enforced sandbox limited to the current workspace, alongside an approval policy. Network access is off by default.

In our coding agent data, 5 to 7% of Claude Code API calls returned a non-200 status in April 2026. Model server errors are a separate measure: in our production rankings for September 7 to November 9, 2026, both Claude Opus 5.5 and GPT-6.1 Sol had a 0.1% 5xx error rate. These rows cover model traffic across Requesty.

Which should you choose?

For per-token billing and routing controls, both tools can connect through a gateway. Claude Code uses ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN for a gateway connection. With a gateway credential active, requests bill through the gateway and do not draw on the Claude subscription allowance.

For Codex, follow Requesty's Codex guide to configure Requesty as the model provider. Requesty supports the Responses API and provides routing controls, including fallback policies and load balancing. Its pricing is 5% on top of model costs.

Our coding tools comparison also covers Cursor and Aider. Before choosing, run your own tasks through the tools you are considering.