Codex CLI vs Claude Code CLI Cost: Plans, API and Cache

Short answer

Start with the CLI whose included allowance covers your chosen model and actual workload. That can minimize additional spending if you already pay for the plan. For API-funded work, compare the model, cache behavior and completed result: neither Codex CLI nor Claude Code has one universal task price.

GPT-6 Astra and Claude Fable 5.1 have equal Standard input/output prices, but Fable 5.1 has cheaper cache reads. GPT-5.6 Sol, Terra, Claude Opus 5 and Sonnet 5 offer lower base rates. The calculations below show how those differences affect a short task and a session that repeatedly reads repository context. These are illustrative budgets, not hands-on test results. OpenAI API pricing, Claude API pricing.

Identify the billing route

Before comparing prices, record the active account, authentication method, workspace or API organization, included allowance, remaining credits and person responsible for the bill. A browser login alone does not prove subscription billing: Claude Code also supports browser-based Console authentication. Codex authentication, Claude Code authentication.

OpenAI distinguishes ChatGPT allowances from API-key pricing. Anthropic likewise separates subscription access from Console billing; its Pro/Max guide warns that an ANTHROPIC_API_KEY can take precedence over subscription login. Check the active route before running a large job. OpenAI billing routes, Claude subscription authentication.

Model entitlement is another boundary. Astra is rolling out to eligible accounts; its availability in your CLI must be confirmed even if you have Codex access. OpenAI says eligible Astra usage draws on existing subscription allowances or workspace billing. Astra availability.

Fable 5 and 5.1 together can use up to 50% of regular weekly allowance on Max and qualifying Premium seats. That is shared capacity, not an additional allowance for each model. Pro and Standard seats use paid usage credits for Fable from the start. Fable plan rules.

Compare current model and cache rates

The following are direct API prices in USD per one million tokens, using Standard processing and OpenAI's short-context tier. Claude write prices use its five-minute cache duration. They are not ChatGPT credit rates, subscription token entitlements or third-party cloud quotes. OpenAI rate card, Anthropic rate card.

Sol's listed rates are promotional; recheck before a longer commitment. Sonnet 5's $2/$10 input/output rates are its Standard prices, following cancellation of the previously announced increase. OpenAI pricing notes, Claude pricing notes.

Context tier and session length are different. Astra requests above 272,000 input tokens charge twice the input/cache rates and 1.5 times the output rate for the full request: $20 uncached input, $2 reads, $25 writes and $75 output per million. Fable 5.1, Opus 5 and Sonnet 5 retain Standard API rates across their supported 1M context window; subscription context access can differ. Astra model pricing, Claude long-context pricing.

Cache lifetime also matters. OpenAI documents at least 30 minutes after a write or reuse for GPT-5.6 and later. Claude's default five-minute lifetime refreshes on use; its one-hour writes cost $20 for Fable 5.1, $10 for Opus 5 and $4 for Sonnet 5 per million tokens. A cheap read rate helps only when reuse succeeds. OpenAI caching, Claude caching.

Keep speed and reasoning modes separate. Astra API Fast is twice the applicable Standard rates, while ChatGPT-authenticated Astra Fast consumes 2.5 times Standard credits. Claude Code Fast is an Opus 5/4.8 research-preview configuration priced at $10 input/$50 output per million; subscription users fund it with usage credits rather than included limits. It is not Fable Fast. First enabling Claude Fast within a conversation also charges its existing context at the Fast uncached-input rate. OpenAI API pricing, Codex speed, Claude Code Fast.

A Pro subscription, Pro reasoning mode and Fast processing are different purchases or settings. OpenAI documents GPT-5.6 Pro reasoning as more aggregate model work billed at the selected model's Standard token rates. Astra Pro has no separately verified rate here; leave its estimate unresolved rather than infer a rate or token multiplier from Standard or Fast. OpenAI reasoning modes.

Two worked workload examples

Use four non-overlapping token buckets. With counts measured in millions, model cost is uncached input × input rate + cache reads × read rate + cache writes × write rate + billable output × output rate.

OpenAI's total input count includes reads and writes, so subtract both to obtain ordinary input. Claude's input_tokens already excludes its separately reported cache reads and writes. Charging cached tokens again as ordinary input overstates either bill. OpenAI usage accounting, Claude usage accounting.

Both examples assume Standard processing, no Pro mode, no regional surcharge or negotiated discount, and zero separately billed hosted-tool or external-service calls. Billable output includes generated reasoning/thinking; do not add it a second time or count only the visible answer. OpenAI reasoning billing, Claude thinking billing.

Short interactive task

Assume a small code explanation or patch uses 8,000 uncached input tokens, zero cache reads, zero cache writes and 2,000 billable output tokens. The no-cache assumption is deliberate; actual CLI sessions can write cache even on their first request.

Astra: 0.008 × $10 + 0 × $1 + 0 × $12.50 + 0.002 × $50 = $0.18. Fable 5.1 also totals $0.18. The same assumed counts cost $0.072 on Sol, $0.040 on Terra, $0.090 on Opus 5 and $0.036 on Sonnet 5. These totals use the Standard rate table above.

The frontier pair ties because this workload gets no cache-read discount. A cheaper model is the sensible first trial if it can produce an accepted result without additional attempts.

Cache-heavy long session

Assume ten requests reuse a 100,000-token repository prefix: one 100,000-token cache write followed by nine successful reads totaling 900,000 tokens. Across the session, add 100,000 uncached input tokens and 10,000 billable output tokens. Total input processed is 1,100,000 tokens, but assume every individual request stays below 150,000 input tokens, so Astra's long-context premium does not apply.

All Claude reuse requests begin within five minutes of the preceding write/read; OpenAI reuse also succeeds. The uncached bucket includes all remaining instructions, history and tool-result text. These are assumed aggregate meters, not predicted token counts for a particular repository.

For example, Fable 5.1 is 0.1 × $10 + 0.9 × $0.25 + 0.1 × $12.50 + 0.01 × $50 = $2.975. Its $0.675 advantage over Astra comes entirely from cache reads under these assumptions. Neither model is guaranteed to use the same number of tokens to finish the same task.

Long pauses, changing prefixes, additional reasoning, retries and tool calls can change the totals. If a required rate or charge is unknown, report the known subtotal and leave the final total unresolved.

When a subscription or API route fits

Use included access first for interactive work when the model is available and your normal sessions fit the allowance. ChatGPT offers limited Codex access on Free/Go and larger paid allowances; Plus is $20 monthly. Claude Pro is $20 month to month or $200 billed annually and includes Claude Code, subject to shared limits. Compare the whole subscription's value rather than dividing its fee by an API token price. ChatGPT plans, Work/Codex pricing, Claude plans.

Choose API billing when you need a separate project budget, attributable usage or repeatable automation. For intermittent work, measured API spend may be lower than an additional subscription; for frequent interactive use, an existing allowance may be more economical. Compare your actual monthly task mix, including overflow. Claude usage credits can extend paid-plan work at Standard API rates, but remain a paid-plan billing route. Codex API access, Claude usage credits.

Try Terra or Sonnet 5 for routine changes, and retain Sol or Opus 5 if they already meet the acceptance bar. Escalate to Astra or Fable 5.1 when a difficult task justifies measuring whether stronger reasoning reduces failed attempts or review repair. OpenAI reports token-efficiency gains for Astra in selected evaluations; Anthropic positions Fable for its hardest, longest tasks. Those are vendor claims, not proof of your repository's cost per accepted change. OpenAI model guidance, Claude Code model guidance.

Before you switch

Run a small comparison on representative tasks: a routine fix, a debugging investigation and a longer refactor. Record the active model/version, effort and speed mode, token categories, extra charges, elapsed time, retries and human review effort. Judge cost per accepted change with passing tests, not cost per generated answer.

Confirm the account and billing owner again after changing credentials, models or providers. Check remaining allowance, credit spending controls, cache duration and per-request context size. Claude's fable alias resolves to Fable 5.1 from version 2.1.255 unless overridden; Fable is not the account-type default. Astra access remains subject to rollout and workspace settings. Claude Code configuration, Astra rollout.