Codex vs Claude Code: Pricing, Limits, and Which Is Cheaper in 2026

Codex vs Claude Code: Pricing, Limits, and Which Is Cheaper in 2026

Codex is the cheaper agent to run and Claude Code is the faster one. In independent testing, the Codex side completed identical tasks for less than half the measured cost, while the Claude Code harness finished the same work in about half the median time. Both agents ride their vendors' chat subscriptions, so $20 to $200 a month buys either one, and both meter usage with hard caps plus an API-key escape hatch for when the caps bite. One fresh wrinkle on the OpenAI side: the $200 Pro tier reopened on September 29 after a three-week pause, cut to ten times the Plus allowance for new sign-ups, and a new $500 tier now sits above it at twenty-five times Plus 1 2. Interactive develop-test-iterate work favors Claude Code. Well-specified volume, GitHub delegation, and cheap throughput favor Codex.

What the same $20, $100, or $200 buys on each side

The tier structures mirror each other to the dollar up to $200, with OpenAI now stacking a $500 Pro tier above the shared ladder. Codex is bundled into ChatGPT plans, Claude Code into Claude plans, and the decision is really about what each seat's usage allowance holds.

Codex side: - Plus, $20 per month: full cloud access. On the current GPT-6 lineup the published estimates run roughly 5 to 45 local messages per five-hour window on GPT-6 Astra, 15 to 160 on the GPT-6.1 Sol that arrived September 29, 15 to 150 on GPT-6 Sol, and 350 to 3,000 on the lightweight GPT-6 Luna, while cloud chats can use more of the allowance than local messages 3. - Pro 100, $100: five times the Plus allowance, the tier OpenAI added on April 9, 2026 in what eesel's pricing guide called its direct response to Claude Code's $100 per month pricing 4. The old five-hour message estimates no longer apply on Pro: OpenAI's pricing guide now says Pro plans carry no five-hour limit at all, only the weekly cap 3. - Pro 200, $200: paused on September 10, 2026 amid unprecedented GPT-6 Astra demand 5 6, then reopened at DevDay on September 29 with the included allowance halved to ten times Plus for new sign-ups 1 2. Subscriptions active between September 22 and 10 a.m. Pacific on September 29 keep the old 20x allowance through October 29, and OpenAI's email to those subscribers adds a one-time grant of 62,500 usage credits, worth $2,500 by OpenAI's own valuation and expiring December 31 1 7. - Pro 500, $500: new at DevDay, twenty-five times the Plus allowance and the only Pro tier with Astra Ultrafast, a speed mode OpenAI rates at up to eight times faster generation, about 300 tokens per second, that also draws the included allowance eight times as fast; Nerd's Chalk works that out to the full 25x budget in Ultrafast lasting only about as far as 3x Plus at standard speed 1 3. - Below Plus sit a limited free tier and an $8 Go plan for light local work 4. Business seats now come in two types: a standard seat at $25 per month or $20 on annual billing, and a $100 Premium seat, announced August 10 and rolled out August 25, that carries five-times-Plus Codex allowances, with workspaces able to buy extra credits 8 9 3. Enterprise and Edu seats on flexible pricing draw on a shared credit pool with no fixed rate limits, while those without flexible pricing get Plus-level per-seat limits 3, and Codex usage now shares one meter with ChatGPT Work 3.

Claude side: - Pro, $20 monthly ($17 per month billed annually, $200 up front), includes Claude Code, worth roughly 45 messages per five-hour window 10 11. - Max 5x, $100: about 225 messages per five-hour window. Max 20x, $200 11. - Team seats run $20 to $100 per seat per month on annual billing, $25 to $125 on monthly billing; Enterprise is $20 per seat plus usage at API rates 10.

Since April 2, 2026, Codex has billed in token credits rather than per message 4. GPT-5.5 still costs 125 credits per million input tokens, 12.5 per million cached, and 750 per million output, but it retires from Codex on all plans on October 14, 2026. The GPT-6 family now sets the meter: Astra at 250 input, 25 cached, and 1,250 output credits per million, GPT-6 Sol at 50, 5, and 250, its September 29 sibling GPT-6.1 Sol matching those input and output rates while halving cached input to 2.5, and Luna at 2.5, 0.25, and 12.5. GPT-5.6 Sol bills 100, 10, and 500 on promotional pricing OpenAI has committed to at least through November 21, 2026, and OpenAI's docs now put a typical GPT-5.6 Sol task at 2 to 15 credits 3. Two ratios still do most of the budgeting work: cached input bills at one tenth of fresh input on nearly every model in the card, with GPT-6.1 Sol the new exception at one twentieth, and output runs five times input across the GPT-6 family and GPT-5.6 Sol, against six times on the older GPT-5.5 and GPT-5.4 lines 3. Anthropic, meanwhile, doubled Claude Code's five-hour limits for Pro, Max, Team, and seat-based Enterprise plans on May 6, 2026, and removed the peak-hours reduction for Pro and Max accounts 12. A week later, on May 13, Anthropic also lifted Claude Code's weekly limits 50 percent above their standard level as a temporary boost, and on August 29 it announced where that landed: the boost ran through September 13, and since September 14 weekly limits sit permanently 25 percent above the pre-boost standard, a level Anthropic's own announcement puts at a 17 percent reduction against the boosted level 13 14 15. The five-hour window was not part of that change, and on September 22, alongside the Opus 5.5 launch, Anthropic raised the five-hour limits again for the same four plan families and gave subscribers a rate-limit reset they can save and spend whenever they choose 15 16. Model choice still dominates the Claude bill, though the Opus premium is now shrinking: Composio measured Opus draining the allocation 5 to 10 times faster than Sonnet, and Opus 5.5 bills 20 percent less per token than Opus 5 at $4 versus $5 per million input and $20 versus $25 per million output, while using fewer tokens per task for a drop Anthropic puts at 40 percent on typical workloads 11 16. The workhorse tier then got cheaper twice: Sonnet 5.5, shipped September 28, runs 30 percent faster and costs up to 30 percent less, landing at $2 per million input and $10 per million output against Sonnet 5's $3 and $15, and Haiku 5.5, shipped October 7, becomes the cheapest model on the card at $0.10 in and $0.50 out per million for prompts under 100,000 tokens 17 10.

The harness is faster on one side, the meter cheaper on the other

Composio, an agent-integrations platform, isolated the two harnesses by running the same model, DeepSeek V4 Flash, through both at max reasoning: 30 identical multi-app tasks, 16 of 30 passed by each, and the same score reached very differently 11. Claude Code finished in a 122.7-second median against Codex's 245.0, using 358 tool calls to Codex's 448, but the entire Codex run cost an estimated $1.294 against Claude Code's $3.116. Per successful task that is $0.081 for Codex versus $0.195 for Claude Code.

Run that arithmetic out to an hour of agent runtime: at those medians Claude Code completes about 29 tasks per hour for roughly $5.72, while Codex completes about 15 for roughly $1.19. Codex delivers half the throughput at about one fifth the hourly cost, and Claude Code delivers roughly double the feedback speed at about two and a half times the per-task cost.

With each vendor's own models in the harness, the same split shows up flipped by reliability. On Composio's 47-scenario Golden Eval, Claude's Fable 5 passed 47 of 47 while GPT-5.6 Sol high passed 45, with Sol using about 19 percent fewer runtime tokens and finishing about 6 percent faster. On the DeepSwe benchmark both hit the same 69 percent pass rate, but Fable High cost an estimated $9.18 per task using 57,000 output tokens and 59 steps, against Sol High's $3.47, 28,000 tokens, and 37 steps 11. That is the whole trade in one line: the reliability premium costs about 2.6 times per task, and the speed premium costs about 2 times in wall-clock patience.

Fourteen dated changes in under seven months outdate the published comparisons

This is why comparison posts go stale so fast on this pair. The dated changes, in order: - April 2, 2026: OpenAI switches Codex from per-message billing to token credits on all plans 4. - April 9, 2026: the $100 Pro 5x tier appears, filling the gap between $20 and $200 4. - May 6, 2026: Anthropic doubles Claude Code's five-hour limits and drops peak-hour throttling 12. - May 13, 2026: Anthropic lifts Claude Code's weekly limits 50 percent above standard as a temporary boost 13 14. - July 12, 2026: OpenAI temporarily removes Codex's five-hour window for Plus, Business, and Pro plans, keeping the weekly credit cap, in a post by its Codex engineering lead that also shipped efficiency changes to GPT-5.6 Sol 18 19. - August 25, 2026: the five-hour limit returns for Plus subscribers, layered back on top of the weekly quota; hitting either one means waiting for a reset or buying extra credits 20 21. - August 29, 2026: Anthropic announces the weekly boost ends September 13 and a permanent level 25 percent above standard begins September 14, which Anthropic puts at 17 percent below the boosted level 13. - September 10, 2026: OpenAI pauses new sign-ups and upgrades to the $200 Pro 20x tier, citing demand for GPT-6 Astra, which launched September 3; existing subscribers keep their plans and no end date is announced 5 6. - September 14, 2026: the transition lands on schedule; the promotion page now reads that weekly limits in Claude Code are 25 percent higher than before the promotion for Pro, Max, Team, and seat-based Enterprise plans 15. - September 22, 2026: OpenAI adds GPT-6 Sol and Luna to Codex 22; the same day Anthropic ships Claude Opus 5.5, cuts its per-token prices 20 percent below Opus 5, and raises Claude Code's five-hour limits again 16. - September 28, 2026: Anthropic ships Sonnet 5.5, which its newsroom bills as 30 percent faster and up to 30 percent cheaper than Sonnet 5; on the API card it lands at $2 per million input and $10 output against Sonnet 5's $3 and $15 17 10. - September 29, 2026: at DevDay OpenAI reopens the paused $200 Pro tier at ten times Plus instead of twenty, adds a $500 Pro tier at twenty-five times, and ships GPT-6.1 Sol; eligible Pro 200 subscribers keep the old allowance through October 29 and receive 62,500 bonus credits 1 2. - October 7, 2026: Anthropic ships Haiku 5.5 with API input at $0.10 per million for prompts under 100,000 tokens, a 90 percent cut from Haiku 4.5's $1 17 10. - October 14, 2026: GPT-5.5, the model whose rates anchored the April credit switch, is set to retire from Codex on all plans 3.

A comparison written before April 2 describes a billing model that no longer exists; Builder.io's matchup was updated April 1, one day before the switch 23. Composio's August post carries limits data "as of June 2026," so even the freshest snapshot predates the restored five-hour cap, the September 14 weekly-limit level, both vendors' September 22 model turnovers, and OpenAI's September 29 Pro restructure 11.

The API escape hatch, and when it beats the subscription

Both agents can run on pay-per-token API billing instead of plan credits. On the Codex side, that track trades away the cloud features, no automatic GitHub code review and no Slack integration, and newer models arrive later than for subscribers 3. On the Claude side, Claude Code meters by API token consumption whenever usage is not covered by a plan, and Anthropic's own cost documentation puts enterprise deployments at around $13 per developer per active day, $150 to $250 per developer per month, with 90 percent of users below $30 per active day 24. The Opus 5.5 cut lands directly on this meter: cache reads, which Anthropic says make up the majority of agentic and coding work costs, drop 60 percent to $0.20 per million tokens, and Sonnet 5.5 pushes cache reads down to $0.10 16 10.

Put Anthropic's documented spend band next to its price list and the subscription logic becomes visible: the $150 to $250 average metered bill sits directly on top of the Max tier's $100 to $200 prices, so a median-to-heavy developer's subscription works as a ceiling on what the same workload would meter to. The OpenAI side lands in the same place: a published OpenAI estimate relayed by eesel puts power users at $100 to $200 per developer per month, exactly the Pro 100 and Pro 200 prices, though the $200 tier has bought half the allowance it did before September 29 4 1. Both vendors anchor their top seats at the metered spend their own materials document, Anthropic's average enterprise bill and OpenAI's power-user estimate.

One tripwire cuts against the escape hatch: a stray ANTHROPIC_API_KEY in your shell silently bills Claude Code at API rates while your subscription sits unused 11. The bypass only saves money if you notice it is on.

Match the agent to the loop, not the benchmark score

Claude Code earns its seat for interactive work: long tool-heavy sessions where its context handling holds up, the skills ecosystem Anthropic built the standard for, and tasks where one wrong action is expensive, the exact use case Composio still hands to Fable despite its higher meter 11. It is the wrong pick if the $20 tier is your budget: Claude Pro "goes through usage fast, sometimes inside the first hour" 11.

Codex earns its seat for delegation and volume: cloud task handoffs, in-loop code review, and the GitHub integration Builder.io's team found genuinely useful on Codex after a weak experience with Claude Code's version 23. At the $20 floor it is the clear value: on the current card, GPT-6 Sol and its new 6.1 sibling bill 50 credits per million input tokens against Astra's 250 and GPT-5.5's 125, and GPT-6.1 Sol's 2.5-credit cached input makes repeat-heavy loops cheaper still, so routing routine work down the lineup stretches the Plus allowance, and Plus users report far more headroom before the wall 3 23. It is the wrong pick if you need the open Agents.md instructions standard, which Claude Code does not support 23.

The pragmatic close: the pair trades wins by release cycle, and the cheapest move for anyone doing both loops is keeping both installed, which is exactly what the tester who ran both for months concluded 11.

Verified October 2026 against Anthropic's pricing, limits, cost, and newsroom pages, OpenAI's Codex pricing documentation and rate card, and Composio's published benchmark runs.