Claude Max vs ChatGPT Pro for coding in October 2026: we priced both from our logs

Updated October 10, 2026 · first published October 10, 2026

Same $100 a month, two very different deals. We priced every Claude Code turn on our Claude Max 5x plan and every Codex request on our ChatGPT Pro $100 plan, at API list prices, from our own logs. Prices changed during the month on both sides, so every figure below is at today's list prices, with the price at the time shown next to it. We use the full Claude weekly cap every week. On the three comparable weeks, one full cap held $658 to $873 of API value, median $761. A full Codex window held a median $179. Claude looks 4x bigger. Then you price the work, and Claude keeps a lead, but a much smaller one.

• Raw API value per full week, today's prices: Claude Max 5x $658 to $873, median $761. Codex Pro $100 median $179.
• Price per token: GPT-6.1 Sol lists at exactly half of Opus 5.5: $2 vs $4 input, $10 vs $20 output.
• Cost per task: Opus 5.5 spent $0.87 on a benchmark task GPT-6 Sol did for $0.26. On Terminal-Bench 4.0, Claude Code with Opus 5.5 spent about 8x what Codex with GPT-6.1 Sol spent for the whole run.
• Tasks per week: about 756 to 1,003 on Claude (median 875), about 688 on Codex. Claude ahead by about 27%.
• The catch on each side: Claude's weekly cap is the ceiling every week, plus a 5-hour limit that stalls long sessions. Codex Pro has no 5-hour limit, and resets can multiply a week.

What changed in October

The last time we ran this comparison, the Claude side looked untouchable: a Max 5x week was worth $1,217 or more at API prices. Three things moved since.

• Anthropic made Claude cheaper per token, so the same quota is worth fewer dollars. Opus 5.5 lists at $4 input and $20 output per million tokens, against $5 and $25 for Opus 5. Its cache reads cost $0.20 instead of $0.50, and a heavy agent session is mostly cache reads.
• The weekly promo ended. From September 14, Claude Code weekly limits carry a permanent 25% bonus, replacing the temporary 50% one. That is about 17% less than the boosted level.
• OpenAI cut prices, then cut Pro $200. GPT-6.1 Sol, live from September 29, halved the cached-input price to $0.10. From October 30, ChatGPT Pro $200 drops from 20x to 10x Plus for Codex. Existing subscribers keep current limits through October 29 and get a one-time $2,500 credit that expires at year end. A new Pro $500 tier is the only one with Ultrafast, and its multiplier is unpublished.

What the vendors claim

Before our numbers, here is what each vendor says you get. These are their claims, not our measurements.

Note what neither publishes: how much a week actually holds. Anthropic gives a per-session multiplier and OpenAI gives task ranges as wide as 10x. That is why we priced our own logs.

How we made the weeks comparable

The four Claude weeks and seven Codex windows were not run under the same prices or the same caps. Priced naively, the oldest weeks look biggest, because the same tokens cost more then. We corrected for three things:

• Model prices. Every Opus 5, Sonnet 5 and Fable 5 turn is repriced at Opus 5.5, Sonnet 5.5 and Fable 5.1 list prices. Every GPT-6 Sol request is repriced at GPT-6.1 Sol. The cache-read rate is the big one on both sides: $0.50 to $0.20 on Opus, $0.20 to $0.10 on Sol.
• Cap size. The Sep 11 Claude week ran under the temporary +50% promo until September 14, then the permanent +25%. It is a different cap from the weeks after it, so we show it but leave it out of the range and median.
• Partial windows. Two Codex windows ended at 98% and 99% of the limit. We scaled them to 100%. A window that stopped at 14% is left out.

Repricing moved the Claude numbers a lot and the headline only a little: unadjusted, the four weeks read $644, $953, $874 and $658. The Sep 18 week fell from $953 to $761, because half of it was Opus 5. On Codex the median moved from $237 at the price of the day to $179 today.

The Claude side: every week is a full cap

We are on Max 5x and we run it to the weekly cap every week. Some weeks we stop at 100% and wait for the reset; one week the client showed the limit message 16 minutes before reset. Either way, each row below is one full weekly cap. Our logs restart on September 11 (older logs were pruned by retention), so we priced four complete windows. Windows start Friday 14:00 UTC, the reset time Claude Code reports.

So one full Max 5x week is worth $658 to $873 of API value at today's prices, median $761. The cap is the same every week; the dollar value is not. The spread comes from what the cap is spent on. Claude.ai chat shares the same pool but is not in Claude Code logs, so a chat-heavy week prices lower. Fable 5.1 lists at 2.5x Opus 5.5 and, per Anthropic, uses limits faster, so its weight against the cap is not its dollar share. We look at that in a separate piece on Fable 5.1 vs Opus 5.5.

Separately, we hit the rolling 5-hour session limit 17 times between September 10 and October 9. That is the real day-to-day friction on Claude: a long session stalls mid-task, and you wait.

Raw data: claude-max-5x-weekly-windows.csv.

The Codex side: smaller windows, cheaper work

Seven full Codex windows on Pro $100, from September 27 to October 15. Overlapping start dates are resets: a reset opens a fresh window before the old one runs out.

At today's prices that is $103 to $237 a window, median $179, under 3x a $67 Plus week rather than the advertised 5x (full breakdown here). Raw data: codex-pro-100-windows-repriced.csv. But Codex does the same job with far fewer dollars:

Snorkel's board is the only one we found that runs each model in its own vendor's agent, Claude Code and Codex, which is what you get on a subscription. Anthropic's own Terminal-Bench figure for Opus 5.5 is 66.4%; OpenAI has not published one for GPT-6.1 Sol and reports DeepSWE instead (75.2%). Our per-task math below uses DataCamp's 3.3x gap, the conservative end: Snorkel's run suggests the gap at max effort is wider. DataCamp's $0.26 was measured on GPT-6 Sol, before the cached price was halved, so on GPT-6.1 Sol the Codex task is, if anything, a little cheaper than we assume.

Pro plans also have no 5-hour limit, and Codex has resets. In 12 days we got eight windows instead of two, worth $1,211 of API value. Resets are not something you can plan around, so we leave them out of the baseline below.

More for your buck: tasks per week

Dollars of API value are the wrong unit when one model needs 3x the tokens for the same job. Divide each plan's weekly value by its cost per task:

One task here means one benchmark-sized coding job. Your tasks will be bigger or smaller, but the ratios hold as long as the 3.3x cost gap does.

Verdict for October 2026

• At $20: ChatGPT Plus. 258 tasks a week is the best unit price on the board. We did not measure Claude Pro, but its weekly pool is a fifth of Max 5x.
• At $100: Claude Max 5x by about 27% on tasks, less if you work at max effort. It does more tasks and wins on hard, ambiguous problems. Codex Pro $100 never stalls on a 5-hour limit, and a reset can triple a week. If your work is a queue of well-specified tickets, take Codex. If it is gnarly debugging and design, take Claude.
• At $120: Claude Max 5x plus ChatGPT Plus is our pick. Route the hard thinking to Opus, the volume to Codex, and you rarely hit either wall.
• At $200: until October 29, ChatGPT Pro $200 is the clear volume winner. From October 30, at our measured ratio, it lands near Claude Max 20x. Lock in the $2,500 credit before then if you are already subscribed.

What this does not show

This is one heavy user's logs, not a panel. Claude logs before September 11 were pruned, so we could not re-check the August weeks behind our earlier $1,217 figure. Claude.ai chat shares the weekly pool but is not in the logs, so a full cap spent only in Claude Code is likely worth more than the $658 low end. Repricing assumes the plan cap did not change when the models did; neither vendor says whether a cheaper model also stretches the cap. A few hundred Codex requests on an older gpt-5.5 model have no published price and are left out. The $67 Plus week was measured earlier and is not repriced. Vendor claims for GPT-6.1 Sol come from launch coverage quoting OpenAI, because OpenAI's own post blocks automated readers; the $2 and $10 prices agree across every source we checked. Dollar figures are API-equivalent value at list price, not money we paid. Both vendors meter in their own units, and neither publishes the conversion.

Sources

• Anthropic API pricing
• Anthropic: What is the Max plan
• OpenAI: Codex plans and usage limits
• The Next Web: OpenAI cuts Pro $200 usage, adds Pro $500
• OpenAI Help: Codex banked resets
• DataCamp: GPT-6 Sol vs Claude Opus 5.5
• codeongrass: Claude Code vs Codex for heavy users
• Snorkel: Terminal-Bench 4.0 leaderboard
• Anthropic: Introducing Claude Opus 5.5
• Anthropic: Claude Fable models on your plan
• Anthropic: Claude Code weekly limits promotion
• Mixed News: GPT-6.1 Sol pricing
• AI Catchup: GPT-6.1 Sol launch, pricing and benchmarks
• OpenAI: Introducing GPT-6.1 Sol

Related

• Fable 5.1 vs Opus 5.5: is 2.5x worth it?
• Claude Max 20x vs ChatGPT Pro
• ChatGPT Pro $100 is really 2.6x Plus
• OpenAI halved ChatGPT Pro $200
• Codex resets turned $40 into $1,211
• Claude Max 20x vs 5x weekly limit
• Claude Opus 5.5 cost model

Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com