Coding agent plan limits and costs: what actually eats your usage

Shared by Pavel · Updated 2026-11-06

As of 2026-11-05: prices, plans and limits were rechecked against the vendor pages that day. GitHub issue threads and the two blog posts were checked only to confirm they load, not re-read. help.openai.com was read in a browser (it returns 403 to automated fetch). Labels: [vendor] = stated on the vendor’s own page; [measured] = measured data; [opinion] = our judgment or our arithmetic on vendor prices. Prices exclude tax. Vendors change plans often, so recheck the linked page before relying on a number.

Tools covered: Claude (Claude Code, claude.ai), OpenAI Codex and ChatGPT, Cursor, GitHub Copilot, Gemini CLI and Antigravity. Not researched: Windsurf, Cline/Roo, Amp, Zed, OpenRouter, Kimi/GLM/MiniMax coding plans.

Quick answers to the questions people keep asking

1. Why did my 5-hour or weekly meter drain faster since October?

No vendor has published an explanation, and neither OpenAI nor Anthropic publishes the conversion from tokens to meter percent or the ratio between session and weekly limits. Whether metering changed in October cannot be checked from vendor pages.

Public issue threads asking about this include https://github.com/openai/codex/issues/58312, https://github.com/openai/codex/issues/56103, https://github.com/anthropics/claude-code/issues/105621, and https://github.com/anthropics/claude-code/issues/106274.

• Claude [vendor]: “Claude Code sends your full conversation with every request, and each time Claude uses tools it sends another request, carrying that batch of tool results.” Cached history is re-read at the cached rate, so a one-line question in a session open all day still draws usage for the whole conversation. Model, effort level, conversation length, tools and connectors all change consumption. Sources: https://code.claude.com/docs/en/costs and https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work.
• Claude [vendor]: agent teams use about 7x more tokens than a standard session when teammates run in plan mode; subagent requests draw on the same usage. Idle sessions can use tokens: session recap and idle compaction, scheduled loops, cross-session messages and goal check-ins resend the full context. Source: https://code.claude.com/docs/en/costs.
• Codex [vendor]: larger projects, long tasks and long sessions “will use significantly more per message”; model, context, reasoning, tool use, retrieval and caching all affect usage, “so prompt length alone isn’t a reliable estimate”. Every MCP server adds context and uses more of your limit; a large AGENTS.md adds context; image generation uses limits about 3-5x faster. Source: https://learn.chatgpt.com/docs/pricing. The page makes no statement on idle or background usage.
• Cursor, Copilot, Gemini CLI and Antigravity [vendor]: Cursor says models drain the pool at different rates; Copilot says conversation length, agentic features and model choice drive usage. [opinion on documentation read] No vendor page found for these four states anything about idle or background drain or an October change.
• [measured, single user report] One Codex user logged a context-window change from 258,400 to 475,000 tokens on 2026-10-11 and input per call roughly doubling: https://github.com/openai/codex/issues/59022. This is one user’s log, not a vendor statement; the docs say nothing about context-window changes.
• [opinion] History is re-sent each call, and cached tokens still count toward tokens processed in Codex reports. A longer context can therefore make each step drain more. This is a plausible mechanism, not a confirmed cause of the October change.

2. Which pool pays: plan limit, credits, or API credits?

Plan allowance is used first on Claude, Codex and Copilot, and for Cursor’s own models; those three then move to credits or a budget if you enable it. Cursor’s third-party models bill from prepaid credit from the first request. Gemini CLI documents nothing at the limit. If ANTHROPIC_API_KEY is set, Claude Code bills the API, not your plan.

• Claude [vendor]: plan limits (5-hour session and weekly; some plans have a separate weekly Fable limit) are used first. At the limit, wait, upgrade, or continue on usage credits (“extra usage”) if enabled, billed at standard API rates, with a monthly spend limit and auto-reload. Credits are used only after included limits are fully used. claude.ai, Claude Code and Claude Desktop draw from the same pool.
• Claude Code [vendor]: if ANTHROPIC_API_KEY is set, Claude Code bills the API instead of the subscription; /logout then /login returns to the plan. Max includes $100/month and Ultra $200/month in API credits for the Claude Platform. The Team plan’s included monthly API credits do not apply to Claude Code or the Claude apps. Sources: https://claude.com/pricing, https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work, https://support.claude.com/en/articles/11145838-using-claude-code-with-your-pro-or-max-plan, https://support.claude.com/en/articles/12429409-extra-usage-for-paid-claude-plans, https://support.claude.com/en/articles/11845131-use-claude-code-with-your-team-or-enterprise-plan.
• OpenAI Codex/ChatGPT [vendor]: the plan allowance first; credits pay for usage after the included limit (Plus and Pro can buy credits). Any user can continue local chats with an API key at API rates, and API-key use is not counted against ChatGPT limits. Local messages and cloud chats share one allowance; cloud tasks may use more. Source: https://learn.chatgpt.com/docs/pricing.
• Cursor [vendor]: one weekly pool for Cursor models, shared across Cursor, Grok Bot and the Grok apps. Third-party models (Premium and above) are billed at API list price from the first request, from prepaid credit that expires (currently 365 days after purchase). At the limit: wait, use prepaid credit or auto top-up, or upgrade; a monthly additional-usage cap applies. Subagents can run a named third-party model while the picker still shows Auto, Grok or Composer, and bill as third-party usage. Sources: https://cursor.com/docs/models-and-pricing and https://cursor.com/help/models-and-usage/usage-limits.
• GitHub Copilot [vendor]: monthly GitHub AI Credits (1 credit = $0.01); base credits are used first, then flex credits. At the limit: upgrade, set a dollar budget for additional usage, or wait for the 1st (reset 00:00 UTC on the 1st of each month; unused credits do not carry over). Business: included credits are pooled across the organisation, not held per user; once spent, the “AI credits paid usage” policy lets usage continue at per-credit rates up to any admin-set budget. Code completions and next edit suggestions cost no credits. Sources: https://docs.github.com/en/copilot/concepts/billing-and-usage/individuals/billing, https://github.com/features/copilot/plans, https://docs.github.com/en/copilot/concepts/billing/usage-based-billing-for-organizations-and-enterprises.
• Gemini CLI [vendor]: daily request limits per user; the page documents no behaviour at the limit and recommends switching to a pay-as-you-go API key or Vertex AI. Source: https://geminicli.com/docs/resources/quota-and-pricing/.
• Undocumented: whether an edge case draws the plan, a cloud-session credit or API credits is sometimes not covered in the vendor’s own help text. Public threads report cases that contradict Claude’s help text (for example https://github.com/anthropics/claude-code/issues/106611); treat those as reports, not confirmed behaviour.

3. The meter says I have quota left, but the tool says limit reached. Why?

Many such reports are bugs, but one documented cause is that more than one limit exists and the bar you are looking at may not be the one that blocked you.

• Claude [vendor]: limits include the 5-hour session, the all-models weekly limit and, on Max and Ultra, a separate weekly Fable limit. Usage credits are spent only after all included limits are fully used, so paid credits can show spend while the all-models weekly bar reads 99% if another bucket is exhausted. This is a plausible explanation for reports like https://github.com/anthropics/claude-code/issues/105305, not proven for that user. Sources: https://support.claude.com/en/articles/12429409-extra-usage-for-paid-claude-plans and https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work. Where to check: Settings > Usage on claude.ai or /usage in Claude Code (documented at https://code.claude.com/docs/en/costs).
• Claude [vendor]: a model-specific Opus or Sonnet limit allows switching to another family; the session or weekly limit is shared across models, so switching models does not help. Source: https://code.claude.com/docs/en/costs.
• Claude Code [vendor]: hitting the 5-hour limit mid-response can show “Usage limit reached · wrapping up”, a small capped allowance that counts toward the weekly limit (Pro once per week; Max can get it each time; verified Ultra does not). Source: https://support.claude.com/en/articles/11145838-using-claude-code-with-your-pro-or-max-plan.
• Copilot [vendor]: in an organisation, a user-level budget or spending limit set by an admin can block use before the pool is spent. A meter over 100% against an “additional budget” is expected on Copilot Business because the included credits are pooled and usage continues as paid usage (https://github.com/microsoft/vscode/issues/345172 asks about exactly this). The docs do not describe the VS Code meter itself, so this explains the billing model, not the bar’s exact rendering.
• Copilot [opinion]: a credit-limit error with the meter showing 13,868.6 of 24,000, while CLI and local mode work for the same user, points to a client bug: https://github.com/microsoft/vscode/issues/344615.
• Codex: reports of limit errors with quota left (examples https://github.com/openai/codex/issues/57655, https://github.com/openai/codex/issues/59222) are mostly account or client bugs; no vendor page explains them.
• Cursor [vendor]: the models-and-pricing page states “Requests are never downgraded in quality or speed” at the limit; they do not explain mismatches between the meter and an error.

4. Will the December merge of ChatGPT and Codex quotas hurt me?

OpenAI says that starting in December, plan usage will have a shared limit across all of ChatGPT and Codex, but has published no numbers for that shared limit: https://openai.com/business/pricing.

The pricing documentation already says “ChatGPT Work and Codex share usage” and “Local messages and cloud chats share your plan's usage allowance,” while saying nothing about a December change: https://learn.chatgpt.com/docs/pricing. Whether this will cost any given user more cannot be determined from published information. A public feedback thread is https://github.com/openai/codex/issues/58723.

5. Where did my usage go, and how do I see it per tool?

• Claude Code and claude.ai [vendor]: /usage in Claude Code shows plan usage bars and a breakdown by skills, subagents, plugins and individual MCP servers (percent of total), recommendations for behaviours behind at least 10 percent of recent usage, and a Loops row for the heaviest /loop or scheduled tasks (Claude Code v2.1.242 or later). Press d or w for the last 24 hours or 7 days. Figures are approximate, computed from local session history on that machine, and exclude other devices and claude.ai.
• Claude [vendor]: also use /context for what fills context, /insights (which spends tokens itself), /usage-credits for credit balance and spend, and Settings > Usage on claude.ai. The Session block’s dollar figure is computed locally at list price and is not relevant for subscription billing. Team and Enterprise have a spend report in organisation analytics.
• Codex/ChatGPT [vendor]: the usage dashboard (Settings > Usage in ChatGPT, or Usage & Billing in the Codex app) shows current limits and reset times; in a Codex CLI session /status shows remaining limits. Sources: https://learn.chatgpt.com/docs/pricing and https://help.openai.com/en/articles/12642688-using-credits-for-flexible-usage-in-chatgpt-personal-plans.
• Cursor [vendor]: Usage at grok.com/dashboard (weekly pool used, next reset, prepaid credit); legacy plans use the Spending tab in the dashboard; editor settings show the same remaining usage. Sources: https://cursor.com/help/models-and-usage/usage-limits and https://cursor.com/docs/models-and-pricing.
• GitHub Copilot [vendor]: the usage dashboard shows the available allowance and what you used. Update the IDE or client to the minimum listed versions (for example VS Code 1.120, Copilot CLI 1.0.48) or the usage display may be inaccurate. Source: https://docs.github.com/en/copilot/concepts/billing-and-usage/individuals/billing.
• Gemini CLI [vendor]: /stats model shows the session’s token usage and limit information; a summary prints on exit. Source: https://geminicli.com/docs/resources/quota-and-pricing/. Antigravity [vendor]: no usage command or dashboard is named on its pricing page (https://antigravity.google/pricing).
• Per-request or past-month spend views: users ask for them on Claude Code and Codex (for example https://github.com/anthropics/claude-code/issues/107299), and no supported surface beyond the above was found in the docs read [opinion].
• Community tools such as ccusage compute dollar values locally at list price; the vendor says this kind of dollar figure is an estimate and not relevant to billing on subscriptions.

6. Do headless or Agent SDK runs cost more of the plan than interactive ones?

Unverified. Anthropic’s help article (https://support.claude.com/en/articles/15036540) says they draw on subscription limits and is silent on any different rate. The article’s 2026-10-07 update adds that Max, Ultra and Team monthly API credits cover Agent SDK and claude -p runs that use an API key from a linked Console organization (read 2026-11-05). The open question is at https://github.com/anthropics/claude-code/issues/107106.

Per-tool limits and prices

All figures [vendor] unless marked, as of 2026-11-05. Recheck each linked page before relying on a number.

Claude (Anthropic)

• Source https://claude.com/pricing. The page gives no numeric allowance (no tokens, no message counts) for any plan, only multiples of Pro.
• Free $0, no Claude Code; limited usage every 5 hours and every week.
• Pro $20/month, or $17/month shown, $200 billed up front. Claude Code included. More usage than Free in each 5-hour and weekly window (no multiplier published). Fable is not included on Pro; it is paid with usage credits.
• Max $100/month (monthly billing only): 5x Pro usage every 5 hours and every week; Fable included up to half your weekly limit; $100/month in API credits for the Claude Platform; priority access at high traffic.
• Ultra $200/month (monthly only): everything in Max; no 5-hour limit for verified accounts; 10x Pro usage every week; $200/month in API credits; Fable up to half the weekly limit.
• Team Standard seat $25 monthly or $20 annual (“more usage than Pro”); Premium seat $125 monthly or $100 annual (“5x more usage than standard seats”). Team is 2 to 150 seats; no numeric caps for either seat type. Enterprise: $20 per seat per month billed annually, 20-seat minimum, plus usage at API rates.
• Team includes “Up to $500 per month in pooled API credits” on claude.com/pricing (seen 2026-11-05); article 11845131 says the monthly API credits included with Team plans can’t be used in Claude Code or the Claude apps, including for extra usage.
• Paid plans have a session limit that resets every five hours (the window starts at your first message) and weekly limits that reset once a week at a fixed time per account. Weekly and monthly caps and model or feature limits may also apply “at our discretion”. Source: https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work.
• API prices (https://platform.claude.com/docs/en/about-claude/pricing), USD per million tokens: input / 5-minute cache write / 1-hour cache write / cache read / output. Fable 5.1: 10 / 12.50 / 20 / 0.25 / 50 (Fable 5: cache read $1). Opus 5.5: 4 / 5 / 8 / 0.20 / 20. Opus 5 and 4.5 to 4.8: 5 / 6.25 / 10 / 0.50 / 25. Sonnet 5.5: 2 / 2.50 / 4 / 0.10 / 10. Sonnet 5: 2 / 2.50 / 4 / 0.20 / 10. Sonnet 4.6 and 4.5: 3 / 3.75 / 6 / 0.30 / 15. Haiku 5.5: 0.10 / 0.125 / 0.20 / 0.01 / 0.50 up to 100k tokens, 0.50 / 0.625 / 1 / 0.05 / 2.50 above 100k. Haiku 4.5: 1 / 1.25 / 2 / 0.10 / 5.
• Batch API: 50 percent off input and output; batch and caching discounts stack. Fast mode (Opus) costs more and is not available in batch. US-only inference is a 1.1x multiplier. Claude 4.7 and later models use a tokenizer that produces about 30 percent more tokens for the same text, so per-token comparisons across generations understate cost differences.
• Caching (https://platform.claude.com/docs/en/build-with-claude/prompt-caching): multipliers on base input are 1.25x for a 5-minute write, 2x for a 1-hour write, 0.1x for a read (0.05x on Opus 5.5 and Sonnet 5.5; 0.025x on Fable 5.1). Default TTL 5 minutes, refreshed free on each hit. Minimum cacheable prompt 512 tokens on the newest models; 1,024 / 2,048 / 4,096 on older ones. The prefix must be byte-identical.
• Claude Code caching (https://code.claude.com/docs/en/prompt-caching): the main conversation gets a 1-hour cache only on a subscription within included usage; with an API key, a cloud provider or usage credits it gets 5 minutes unless promptCacheTtl or CLAUDE_CODE_PROMPT_CACHE_TTL is set to 1h. Subagents, workflows, teammates and compaction get 5 minutes by default. Each model has its own cache: switching model mid-session re-reads the whole history uncached; compaction, MCP server changes (when tool search is off), fast-mode toggles and Claude Code upgrades also invalidate it.
• [measured, vendor] Across enterprise deployments, about $13 per developer per active day and $150-250 per developer per month, below $30 per active day for 90 percent of users (https://code.claude.com/docs/en/costs). Method and sample size are not stated.

OpenAI Codex and ChatGPT

• Plan prices: Free $0, Go $8, Plus $20, Pro from $100 (Pro tiers $100, $200 or $500 per month). Sources: https://chatgpt.com/pricing and https://learn.chatgpt.com/docs/pricing.
• Local-message estimates per 5-hour window, Plus and Standard Business: GPT-6 Astra 5-45; GPT-6.1 Sol 15-160; GPT-6.1 Terra 35-285; GPT-6 Sol 15-150; GPT-6 Luna 350-3,000. The page says these “are not fixed message limits”. “Pro plans currently have no five-hour limit”; weekly limits “may also apply”, with no figures published.
• Business (https://openai.com/business/pricing): Standard seat $20 per month billed annually or $25 monthly; Premium seat $100 annually or $125 monthly, “no 5-hour limit”; 2 to 200 users.
• Credits pay for usage after the included limit. Rates per 1M tokens (input / cached input / output, in credits): GPT-6 Astra 250 / 25 / 1,250; GPT-6.1 Sol 50 / 2.5 / 250; GPT-6.1 Terra 25 / 1.25 / 125; GPT-6 Sol 50 / 5 / 250; GPT-6 Luna 2.5 / 0.25 / 12.5; GPT-5.6 Sol 100 / 10 / 500. A typical GPT-5.6 Sol task “may use 2-15 credits”. No separate cache-write charge in Codex credit billing. Fast mode: 2.5x included usage, 2x purchased credits. Ultrafast (GPT-6 Astra and GPT-6.1 Sol only): 8x included, 6x purchased. GPT-5.6 Sol promotional pricing runs “at least through November 21, 2026” on the Codex page and until November 30, 2026 on the API page. Credits are valid 12 months from purchase and non-refundable unless required by law or contract or an approved exception applies (help.openai.com article 12642688, read in a browser 2026-11-05; help.openai.com returns 403 to automated fetch, and learn.chatgpt.com/docs/pricing does not state credit validity).
• The credit-to-dollar rate is not stated on any vendor page read. [opinion, derived from vendor figures] 1 credit = $0.04, because every credit rate is 25x the API price of the same model (GPT-6.1 Sol $2 / $0.10 / $10 per MTok versus 50 / 2.5 / 250 credits) and desktop voice is $0.05 per minute for 1.25 credits. This is our derivation, not an OpenAI statement.
• API prices (https://developers.openai.com/api/docs/pricing), USD per 1M tokens, input / cached input / output, standard speed: gpt-6-astra $10 / $1 / $50; gpt-6.1-sol $2 / $0.10 / $10; gpt-6.1-terra $1 / $0.05 / $5; gpt-6-luna $0.10 / $0.01 / $0.50; gpt-5.6-sol (promotional until 2026-11-30) $4 / $0.40 / $20; gpt-5.3-codex $1.75 / $0.175 / $14. Regional processing (data residency) carries a 10 percent uplift for eligible models released on or after 2026-03-05. Flex/Batch, Fast and Ultrafast prices exist but were not captured.

Cursor

• Sources https://cursor.com/pricing and https://cursor.com/docs/models-and-pricing. Cursor is now sold as “The X Subscription, one subscription for Grok, Cursor, and X”; the weekly pool is shared across Grok and Cursor. Older plans (Hobby, Start, Pro, Pro+, Ultra, and Teams Standard and Premium seats) are “legacy Cursor plans” that keep working with two monthly pools and on-demand usage.
• Tiers: Free $0 (limited); Plus $8; Premium $30 (4x Plus); Super $100 (15x Plus, priority access at peak); Ultra $200 (50x Plus). Annual: Plus $80, Premium $300, Super $1,000, Ultra $2,000. Teams $30, $100 or $200 per seat per month (Plus not available). Enterprise is custom.
• One weekly pool for Cursor models, resetting weekly; no dollar or token size of the pool is published. Plus has no Auto, Cloud Agents, Bugbot, third-party models or own API keys.
• Sample overage rates per MTok (input / cache read / output): Grok 4.6 $2 / $0.5 / $6; Composer 2.5 $0.5 / $0.2 / $2.5. On Teams and Enterprise, third-party models carry an extra “Cursor Token Rate” of $0.25 per million tokens.

GitHub Copilot

• Sources https://github.com/features/copilot/plans, https://docs.github.com/en/copilot/get-started/plans and https://docs.github.com/en/copilot/concepts/billing-and-usage/individuals/billing. Billing is GitHub AI Credits, 1 credit = $0.01. Pro $10: 1,000 base + 500 flex = 1,500 credits. Pro+ $39: 3,900 + 3,100 = 7,000. Max $100: 10,000 + 10,000 = 20,000. Business $19 per granted seat: 1,900 credits per user (shared pool); Enterprise $39: 3,900. Base credits match the subscription price and never change; the flex allotment “designed to adapt as the economics of AI evolve”.
• [unverified: not found on the cited Copilot pages on 2026-11-05] Premium-request billing applied until 2026-06-01; it remains only for Pro and Pro+ subscribers on an existing annual plan who stayed on legacy billing (Pro 300, Pro+ 1,500 premium requests, $0.04 each overage).
• Code completions and next edit suggestions are unlimited and cost no credits. Free: 2,000 completions per month. Discounts: 10 percent off model costs with auto model selection; 30 percent off with Copilot Fusion through 2026-11-30.

Gemini CLI and Google

• Gemini CLI daily request limits per user (https://geminicli.com/docs/resources/quota-and-pricing/, page “Last updated: Jun 18, 2026”): Code Assist Individual 1,000; Google AI Pro 1,500; Google AI Ultra 2,000; Workspace Code Assist Standard 1,500, Enterprise 2,000; Workspace AI Ultra 2,000; Gemini API key free tier 250 per day, Flash only. Per-minute limits are mentioned but not quantified. Google AI Plus and Workspace Standard/Plus/AI Expanded are not supported for Gemini CLI.
• Gemini CLI [vendor]: the quota page says Gemini CLI was replaced by Antigravity CLI on June 18th, 2026 for unpaid-tier and Google One users, and that Gemini CLI has not been optimized for Gemini 4 Argon (to use it, the page says to use Antigravity) (https://geminicli.com/docs/resources/quota-and-pricing/, read 2026-11-05).
• Antigravity (https://antigravity.google/pricing and https://antigravity.google/docs/plans, read 2026-11-05): limits are qualitative only. Individual $0 “Basic weekly rate limits”; Google AI Pro “More rate limits” plus a “Flexible AI credit pool”; Google AI Ultra “Even more rate limits” plus the credit pool. No numbers. The docs say Pro and Ultra quota refreshes every five hours until the weekly limit, the free tier refreshes weekly, limits track the work the agent does per prompt, and Pro and Ultra can spend purchased AI credits above the quota (setting: Never or Always). Business via Google Cloud from $30 per seat per month (Gemini Enterprise Standard/Plus), or pay-as-you-go with $0 seat fee.
• Google AI plans (https://one.google.com/about/google-ai-plans/, read 2026-11-05 with the page set to United States, prices in USD): Plus $4.99/month (400 GB); Pro $19.99/month (5 TB; $199.99 billed annually per the savings calculator); Ultra 5x from $99.99/month (20 TB); Ultra 20x $199.99/month (30 TB). Gemini usage limits are 2x (Plus) and 4x (Pro) of non-AI plans, and 5x / 20x of Pro (Ultra). Antigravity agent requests are listed as Limited (Plus), Expanded (Pro), Higher (Ultra 5x) and Highest (Ultra 20x), with no numbers; Gemini 4 Argon access in Antigravity is on Pro and Ultra. Other countries and student offers were not read.
• Gemini API prices (https://aistudio.google.com/docs/pricing, USD per 1M tokens, paid tier Standard tab, input / output / cached input, re-read 2026-11-05): Gemini 4 Argon $2 through 2026-11-30, then $4 from 2026-12-01 / $10, then $20 / $0.10, then $0.20 (not on the free tier); Gemini 3.1 Pro Preview $2 (prompts up to 200k) or $4 / $12 or $18 / $0.20 or $0.40; Gemini 3.8 and 3.6 Flash $0.75 through 2026-12-31, then $1.50 from 2027-01-01 / $3.75, then $7.50 / $0.075, then $0.15; Gemini 3.5 Flash-Lite $0.30 / $2.50 / $0.03; Gemini 3.1 Flash-Lite $0.25 for text, image and video ($0.50 audio) / $1.50 / $0.025. The Batch, Flex and Priority tabs were not read.

Worked arithmetic

Our arithmetic on the vendor list prices above [opinion], not vendor statements.

• Break-even rule: a subscription beats the API when your monthly API-equivalent spend would exceed the plan price AND you stay inside the plan’s windows. Pro ($20) breaks even at $20 of API-rate usage; Max ($100) at $100; Ultra ($200) at $200.
• Worked example: one developer at the enterprise average of $200/month API-equivalent (within the [measured, vendor] $150-250 range). API cost $200; Max ($100) is half; Pro ($20) is cheaper but likely to hit 5-hour and weekly limits; Ultra ($200) only matches. For five developers running many parallel agents (agent teams use about 7x tokens when teammates run in plan mode, per the vendor), API-equivalent spend could pass $1,000/month per person, where Ultra or API with caching decides. Team Premium seats at $125 monthly or $100 annual per seat are the pooled alternative.
• Caching: a 100k-token context, 50 turns, Sonnet 5.5. Uncached input: 50 x 100k x $2/MTok = $10. Cached (one 5-minute write at $2.50/MTok = $0.25, then 49 reads at $0.10/MTok = $0.49): about $0.74 for input, a 93 percent reduction, assuming every turn lands within the TTL and nothing invalidates the prefix. Output tokens are unchanged.
• Model routing: for 1M input + 200k output tokens, Opus 5.5 costs $4 + $4 = $8; Sonnet 5.5 $2 + $2 = $4; Haiku 5.5 about $0.10 + $0.10 = $0.20. The tokenizer difference and capability differences are not priced in.
• Caveat on “what is my plan worth”: figures quoted online (for example “Max 20x is worth about $8,000 of API usage”) are community-reported from log parsing, not published by Anthropic. Break-even thresholds in two blog posts range from about $80 to $100 (https://botfarm-site.pages.dev/blog/claude-max-true-price and https://dev.to/chenyanchen/is-claude-max-worth-it-for-developers-a-real-cost-analysis-64e, seen only in search results on 2026-11-04, not fetched). No two sources agree and none are vendor data.

Ways to stretch a limit

Documented unless marked; no vendor-measured percentage savings for these habits were found, beyond arithmetic from list prices.

• [vendor] Keep sessions on one model and one effort level; model switches and tool-set changes invalidate the cache (https://code.claude.com/docs/en/prompt-caching).
• [vendor] Clear between tasks; compact at natural breaks; avoid resuming old sessions after the cache expires. /clear costs nothing while /compact is itself a large request (https://code.claude.com/docs/en/costs).
• [vendor] Use cheaper models for subagents; keep agent teams small; shut idle teammates down.
• [vendor] Set a 1-hour cache TTL for gaps longer than five minutes when on API or credits.
• [vendor] Trim CLAUDE.md (under 200 lines), defer MCP tool loading, prefer CLI tools over MCP, filter verbose command output with hooks, use specific prompts and plan mode.
• [vendor] API side: batch non-urgent work (50 percent off), pre-warm caches, keep a stable prefix order.
• [vendor] Cursor: Auto “Cost” mode and Cursor-native models draw only from the weekly pool; third-party models cost list price from credit. [opinion] Codex: GPT-6 Luna’s band (350-3,000 per 5 hours) is the largest listed.

What is unverified

Stated as unverified; do not treat any of these as fact.

• Claude: any numeric allowance per plan (not published). Pricing, support articles and platform pages were re-read in a browser on 2026-11-05.
• Whether any vendor changed metering in October: no vendor statement found. Why meters drain faster: no vendor explanation published.
• OpenAI: the credit-to-dollar rate (our derivation only); weekly Codex limit figures (none published); whether Pro keeps “no five-hour limit” beyond today’s wording; what the December shared ChatGPT and Codex limit will be.
• Cursor: pool size in dollars (not published).
• GitHub: credit burn per model; whether Copilot Pro includes third-party agents (the plans page contradicts itself: a bullet says access to Claude Code and Codex, a table row says not included for Pro).
• Google: Antigravity numeric limits (not published); Gemini API Batch, Flex and Priority prices (not read).
• Idle or background usage for Codex, Cursor, Copilot, Gemini CLI and Antigravity: no vendor statement found.
• Whether headless, claude -p or Agent SDK runs consume the plan at a different rate than interactive use.
• Any independent measurement of subscription-to-API value.
• Other tools not researched: Windsurf, Cline/Roo, Amp, Zed, OpenRouter, Kimi/GLM/MiniMax coding plans.

Weekly recheck list

Claude pricing and limit articles; OpenAI pricing page (5-hour bands, credit rates, the December shared limit, GPT-5.6 Sol promo ending 2026-11-21 on the Codex page and 2026-11-30 on the API page); Cursor models-and-pricing; Copilot credit allowances (the flex allotment may change) and the Fusion discount ending 2026-11-30; Google quota page and scheduled Gemini price changes on 2026-12-01 and 2027-01-01; the Claude Code cost averages.