Electricity Bench

Plan and limit changes

Every plan or limit change the coding-agent subscriptions made in 2026, newest first, with what the bench measured on the affected agents in the last weekly run before the change and the first one after it.

The measured line is each affected agent's share of its own plan's weekly limit for one full bench pass (every suite it graded that week, added up, the rule the Overall view uses), read from the published history. Grading started in W32 2026, so changes before August have no run on the near side. A change with no run after it yet says so. 24 of 29 entries link a source that was fetched and confirms the claim; the rest wear an unverified chip until they do. Announcement dates are the provider's; the runs are ours.

September 2026

  1. AnthropicLimit up

    Weekly limits in Claude Code permanently 25% higher than before the promotion

    The permanent increase replaces the +50% promotion that ended the day before, so a subscriber who was on the promo has about 17% less weekly capacity than last week. 5-hour limits are unchanged.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no run after this change yet; last run before it: 6.2% of the weekly limit per bench pass (W37 2026); the W38 2026 run straddles the change
    • Claude Haiku 4.5 on Claude Max 20x: 1.2% to 1% of the weekly limit per bench pass (3 suites, W37 2026 to W38 2026), 17% less
    • Claude Opus 5 on Claude Max 20x: no run after this change yet; last run before it: 2% of the weekly limit per bench pass (W37 2026); the W38 2026 run straddles the change
    • Claude Sonnet 5 on Claude Max 20x: 1.2% to 1% of the weekly limit per bench pass (3 suites, W37 2026 to W38 2026), 17% less

    Source: Claude Help Center: Claude Code May-August 2026 weekly limits promotion

  2. AnthropicLimit down

    The +50% weekly limits promotion in Claude Code ends

    The boost that ran since May 13 for Pro, Max, Team and seat-based Enterprise plans expires at the end of the day.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: 6.2% to 7% of the weekly limit per bench pass (3 suites, W37 2026 to W38 2026), 13% more
    • Claude Haiku 4.5 on Claude Max 20x: 1.2% to 1% of the weekly limit per bench pass (3 suites, W37 2026 to W38 2026), 17% less
    • Claude Opus 5 on Claude Max 20x: 2% to 2.2% of the weekly limit per bench pass (3 suites, W37 2026 to W38 2026), 10% more
    • Claude Sonnet 5 on Claude Max 20x: 1.2% to 1% of the weekly limit per bench pass (3 suites, W37 2026 to W38 2026), 17% less

    Source: Claude Help Center: Claude Code May-August 2026 weekly limits promotion

  3. xAIProductunverified

    Grok 4.7 ships

    The next Grok version lands on the same weekly compute pool. The bench still grades Grok 4.6; 4.7 is a succession candidate.

    Plans: SuperGrok

    Measured
    • Grok 4.6 on SuperGrok: 9.0% to 18.0% of the weekly limit per bench pass (3 suites, W37 2026 to W38 2026), 99% more

    Source: xAI news

  4. OpenAIPricing

    ChatGPT Pro $200 new sign-ups and upgrades paused

    New Pro subscriptions were paused during the GPT-6 Astra rollout. Existing Pro subscribers were not affected.

    Plans: ChatGPT Pro

    Measured

    No graded agent is on one of these plans.

    Source: codexusage.dev: Codex limit changes

  5. OpenAILimit up

    Global usage reset for paid ChatGPT plans, after banked resets on Sep 3 and 4

    OpenAI reset the usage meters of eligible Plus, Pro and Business subscriptions during the Astra rollout, on top of banked resets granted in two waves. Reports of shifted reset timestamps followed on Sep 9 to 11 and are unconfirmed.

    Plans: ChatGPT Plus, ChatGPT Pro, ChatGPT Pro 5x

    Measured
    • GPT-5.6 Luna on ChatGPT Pro 5x: 0.2% to 1% of the weekly limit per bench pass (3 suites, W36 2026 to W38 2026), 370% more
    • GPT-5.6 Sol on ChatGPT Pro 5x: 3.2% to 3% of the weekly limit per bench pass (3 suites, W36 2026 to W37 2026), 7% less
    • GPT-5.6 Terra on ChatGPT Pro 5x: 2.3% to 2.2% of the weekly limit per bench pass (3 suites, W36 2026 to W37 2026), 6% less
    • GPT-6 Astra on ChatGPT Pro 5x: 4.2% to 5% of the weekly limit per bench pass (3 suites, W36 2026 to W37 2026), 19% more

    Source: codexusage.dev: Codex limit changes

  6. OpenAIProduct

    GPT-6 Astra released to paid ChatGPT users

    A limited preview opened on Sep 3 and paid users got the model the next day. It draws from the same Codex allowance as the GPT-5.6 models.

    Plans: ChatGPT Plus, ChatGPT Pro, ChatGPT Pro 5x

    Measured
    • GPT-5.6 Luna on ChatGPT Pro 5x: 0.2% to <1% of the weekly limit per bench pass (3 suites, W36 2026 to W37 2026)
    • GPT-5.6 Sol on ChatGPT Pro 5x: 3.2% to 3% of the weekly limit per bench pass (3 suites, W36 2026 to W37 2026), 7% less
    • GPT-5.6 Terra on ChatGPT Pro 5x: 2.3% to 2.2% of the weekly limit per bench pass (3 suites, W36 2026 to W37 2026), 6% less
    • GPT-6 Astra on ChatGPT Pro 5x: no measured run before this change; first run after it: 4.2% of the weekly limit per bench pass (W36 2026)

    Source: Wikipedia: GPT-6

August 2026

  1. AnthropicLimit down

    Anthropic announces the end of the +50% promo and a permanent +25% from Sep 14

    The ClaudeDevs announcement framed it as a permanent raise over pre-promo limits. Measured against what subscribers had during the promo it is a cut of about 17%.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: 7% to 7.2% of the weekly limit per bench pass (3 suites, W35 2026 to W36 2026), about the same
    • Claude Haiku 4.5 on Claude Max 20x: 1% to 1% of the weekly limit per bench pass (3 suites, W35 2026 to W36 2026), about the same
    • Claude Opus 5 on Claude Max 20x: 2.2% to 2% of the weekly limit per bench pass (3 suites, W35 2026 to W36 2026), 9% less
    • Claude Sonnet 5 on Claude Max 20x: 1% to 1% of the weekly limit per bench pass (3 suites, W35 2026 to W36 2026), about the same

    Source: AI Catchup: Claude Code limits, 25% permanent increase starting September 14

  2. CursorPricing

    Cursor retires the flat Auto rate; Auto bills at the routed model's list price

    Auto requests used to cost a flat $1.25 per million input and $6 per million output tokens whatever model answered. They now bill at the list price of the model each request is routed to, and Teams and Enterprise pay a $0.25 per million token Cursor rate on top for third-party models.

    Plans: Cursor Pro, Cursor Pro+, Cursor Ultra

    Measured
    • Composer 2.5 on Cursor Pro: graded, but this plan is not metered on a weekly window, so there is no weekly figure to compare

    Source: Cursor docs: Models and pricing

  3. OpenAILimit up

    Business Premium seats launch with the 5-hour limit removed

    A new seat tier for Business workspaces meters only on the weekly allowance. Plus and Pro keep both windows.

    Plans: ChatGPT Business

    Measured

    No graded agent is on one of these plans.

    Source: codexusage.dev: Codex limit changes

  4. xAIProduct

    Grok 4.6 ships

    Released through the xAI API, GitHub Copilot and Amazon Bedrock, and picked up by grok-cli. The bench registered it as the successor of Grok 4.5 the same week.

    Plans: SuperGrok

    Measured
    • Grok 4.6 on SuperGrok: 8.0% to 15.6% of the weekly limit per bench pass (real-world suite, W32 2026 to W33 2026), 96% more

    Source: Wikipedia: Grok (chatbot)

July 2026

  1. OpenAIPricing

    GPT-5.6 Luna price cut 80%, Terra 20%

    API and credit pricing for Luna dropped by 80% and Terra by 20%, so the same task consumes less of a credit-metered allowance on those models. Sol followed with a 20% API cut on Aug 21.

    Plans: ChatGPT Plus, ChatGPT Pro, ChatGPT Pro 5x

    Measured
    • GPT-5.6 Luna on ChatGPT Pro 5x: no measured run before this change; first run after it: 0.1% of the weekly limit per bench pass (W32 2026)
    • GPT-5.6 Sol on ChatGPT Pro 5x: no measured run before this change; first run after it: 2.3% of the weekly limit per bench pass (W32 2026)
    • GPT-5.6 Terra on ChatGPT Pro 5x: no measured run before this change; first run after it: 0.9% of the weekly limit per bench pass (W32 2026)
    • GPT-6 Astra on ChatGPT Pro 5x: no measured run before this change; first run after it: 4.2% of the weekly limit per bench pass (W36 2026)

    Source: codexusage.dev: Codex limit changes

  2. Moonshot AIProduct

    Kimi paid memberships sold out; new sign-ups paused after the K3 launch

    Moonshot paused new Kimi and Kimi Code memberships within 48 hours of the K3 launch to protect existing subscribers, promising to reopen spots in batches. Existing plans kept their weekly allowance and 5-hour limit.

    Plans: Kimi Moderato, Kimi Allegretto, Kimi Allegro, Kimi Vivace

    Measured
    • Kimi K3 256K on Kimi Allegretto: no measured run before this change; first run after it: 32% of the weekly limit per bench pass (W34 2026)

    Source: The Vibe Father: Kimi sold out, coding plans paused

  3. AnthropicPromounverified

    The +50% weekly promo is extended past its first end date

    Announced as running through July 13, the boost was extended to July 19, then August 19, then August 31, before the Sep 13 end was set.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no measured run before this change; first run after it: 4% of the weekly limit per bench pass (W32 2026)
    • Claude Haiku 4.5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)
    • Claude Opus 5 on Claude Max 20x: no measured run before this change; first run after it: 1.5% of the weekly limit per bench pass (W32 2026)
    • Claude Sonnet 5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)

    Source: Pasquale Pillitteri: Claude Code weekly limits +50%

June 2026

  1. CursorPricing

    Cursor Teams seats split usage into two pools and add a $120 Premium seat

    The Standard Teams seat now draws from separate Cursor-model and third-party-model pools, and a Premium seat at $120 per month offers five times the Standard usage. Renewing customers moved on July 1.

    Plans: Cursor Teams

    Measured

    No graded agent is on one of these plans.

    Source: Finout: What happened to Cursor pricing, 2026 guide

  2. GoogleProduct

    Gemini CLI and Code Assist stop serving Google AI Pro, Ultra and free users; Antigravity CLI takes over

    Announced on May 19 with the Antigravity CLI launch: from June 18 the Gemini CLI and the Code Assist IDE extensions stop serving requests for AI Pro, AI Ultra and free accounts, which move to the Antigravity CLI on the plan's own limits.

    Plans: Google AI Pro, Google AI Ultra 5x, Google AI Ultra 20x

    Measured
    • Gemini 3.1 Pro on Google AI Pro: no measured run before this change; first run after it: 16.8% of the weekly limit per bench pass (W32 2026)
    • Gemini 3.8 Flash on Google AI Pro: no measured run before this change; first run after it: 19.6% of the weekly limit per bench pass (W32 2026)

    Source: Google for Developers: An important update, transitioning Gemini CLI to Antigravity CLI

  3. AnthropicProduct

    Kahn v. Anthropic: class action filed over the Max 5x and 20x usage claims

    A Max 20x subscriber alleges the delivered usage is far below the advertised multiples, citing 15% of a weekly allowance spent in one 5-hour session. The suit seeks to represent US Max subscribers since April 2025.

    Plans: Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no measured run before this change; first run after it: 4% of the weekly limit per bench pass (W32 2026)
    • Claude Haiku 4.5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)
    • Claude Opus 5 on Claude Max 20x: no measured run before this change; first run after it: 1.5% of the weekly limit per bench pass (W32 2026)
    • Claude Sonnet 5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)

    Source: Engadget: Anthropic hit with lawsuit over its Claude Max usage limits

  4. OpenAIMetering

    Codex rate-limit reset banking, with one free reset for Plus and Pro

    Subscribers can bank a reset of their usage window and spend it when they hit the limit. Promotional resets granted to Plus and Pro expire after 30 days.

    Plans: ChatGPT Plus, ChatGPT Pro, ChatGPT Pro 5x

    Measured
    • GPT-5.6 Luna on ChatGPT Pro 5x: no measured run before this change; first run after it: 0.1% of the weekly limit per bench pass (W32 2026)
    • GPT-5.6 Sol on ChatGPT Pro 5x: no measured run before this change; first run after it: 2.3% of the weekly limit per bench pass (W32 2026)
    • GPT-5.6 Terra on ChatGPT Pro 5x: no measured run before this change; first run after it: 0.9% of the weekly limit per bench pass (W32 2026)
    • GPT-6 Astra on ChatGPT Pro 5x: no measured run before this change; first run after it: 4.2% of the weekly limit per bench pass (W36 2026)

    Source: codexusage.dev: Codex limit changes

  5. xAIMetering

    SuperGrok moves from per-feature daily counts to one weekly compute pool shown as a percentage

    Text chat, Imagine, Voice, Build and API usage now drain the same weekly meter. xAI published no per-tier figures. The change landed in June 2026; xAI did not date it, so this entry sits at the start of the month.

    Plans: SuperGrok, SuperGrok Heavy

    Measured
    • Grok 4.6 on SuperGrok: no measured run before this change; first run after it: 8.0% of the weekly limit per bench pass (W32 2026)

    Source: Robo Rhythms: Grok now counts your text chats in one weekly usage limit

May 2026

  1. OpenAILimit down

    The Codex Pro 2x capacity promotion expires; weekly limits fall further than expected

    Pro subscribers reported the weekly allowance dropping to about a third of the promo level rather than half, and quota draining while Codex was idle.

    Plans: ChatGPT Pro, ChatGPT Pro 5x

    Measured
    • GPT-5.6 Luna on ChatGPT Pro 5x: no measured run before this change; first run after it: 0.1% of the weekly limit per bench pass (W32 2026)
    • GPT-5.6 Sol on ChatGPT Pro 5x: no measured run before this change; first run after it: 2.3% of the weekly limit per bench pass (W32 2026)
    • GPT-5.6 Terra on ChatGPT Pro 5x: no measured run before this change; first run after it: 0.9% of the weekly limit per bench pass (W32 2026)
    • GPT-6 Astra on ChatGPT Pro 5x: no measured run before this change; first run after it: 4.2% of the weekly limit per bench pass (W36 2026)

    Source: openai/codex #26512: Pro 5x weekly limit dropped after June 1

  2. GooglePricing

    Antigravity plan changes: a $100 Ultra at 5x, the top Ultra cut to $200 at 20x, Flash and Pro quotas merged

    Flash and Pro no longer have separate limits; one rate limit is drawn down at API prices. Base plans lose their included AI credits, which become overage only.

    Plans: Google AI Pro, Google AI Ultra 5x, Google AI Ultra 20x

    Measured
    • Gemini 3.1 Pro on Google AI Pro: no measured run before this change; first run after it: 16.8% of the weekly limit per bench pass (W32 2026)
    • Gemini 3.8 Flash on Google AI Pro: no measured run before this change; first run after it: 19.6% of the weekly limit per bench pass (W32 2026)

    Source: Google Antigravity blog: Changes to Antigravity plans

  3. AnthropicPromo

    Weekly limits in Claude Code raised 50% as a promotion

    Pro, Max, Team and seat-based Enterprise plans got 50% more weekly usage in Claude Code, everywhere it runs. 5-hour limits were not affected.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no measured run before this change; first run after it: 4% of the weekly limit per bench pass (W32 2026)
    • Claude Haiku 4.5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)
    • Claude Opus 5 on Claude Max 20x: no measured run before this change; first run after it: 1.5% of the weekly limit per bench pass (W32 2026)
    • Claude Sonnet 5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)

    Source: Claude Help Center: Claude Code May-August 2026 weekly limits promotion

  4. AnthropicLimit up

    Claude Code 5-hour limits doubled; the peak-hours reduction removed

    Anthropic doubled the 5-hour rate limits for Pro, Max, Team and seat-based Enterprise plans and dropped the peak-hour throttling for Pro and Max, citing new SpaceX Colossus compute. Weekly limits were not mentioned.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no measured run before this change; first run after it: 4% of the weekly limit per bench pass (W32 2026)
    • Claude Haiku 4.5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)
    • Claude Opus 5 on Claude Max 20x: no measured run before this change; first run after it: 1.5% of the weekly limit per bench pass (W32 2026)
    • Claude Sonnet 5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)

    Source: Anthropic: Higher usage limits and a SpaceX compute deal

April 2026

  1. OpenAILimit down

    New Codex limit system: usage priced by reasoning time, Business drains far faster

    Subscribers worked out that a Plus 5-hour window bought roughly 40 minutes of GPT-5.4 reasoning and a Business seat about 12 minutes, while Pro 5x ran a 10x Plus allowance during its promotion.

    Plans: ChatGPT Plus, ChatGPT Pro, ChatGPT Business

    Measured

    No graded agent is on one of these plans.

    Source: OpenAI community: Understanding the new Codex limit system after the April 9 update

  2. OpenAIMetering

    Codex moves Plus, Pro and Business from per-message to token-based credit billing

    A message no longer costs one unit. Input, cached input and output tokens are metered separately, so a small fix costs far less of the allowance than a full-repo refactor.

    Plans: ChatGPT Plus, ChatGPT Pro, ChatGPT Business

    Measured

    No graded agent is on one of these plans.

    Source: CloudZero: OpenAI Codex pricing in 2026

March 2026

  1. GooglePricingunverified

    Gemini Pro models become paid-only; the free tier is limited to Flash

    Free users lost access to the Pro models in the Gemini app, which moved behind AI Pro.

    Plans: Google AI Pro

    Measured
    • Gemini 3.1 Pro on Google AI Pro: no measured run before this change; first run after it: 16.8% of the weekly limit per bench pass (W32 2026)
    • Gemini 3.8 Flash on Google AI Pro: no measured run before this change; first run after it: 19.6% of the weekly limit per bench pass (W32 2026)

    Source: Gemini app release notes

  2. AnthropicLimit down

    Abnormal usage drain across paid Claude plans, with peak-hour throttling confirmed

    From March 23 subscribers reported 5-hour sessions emptying in minutes. Anthropic confirmed faster peak-hour drain on March 26; the issue also traced prompt-caching and session-resume bugs and the end of the off-peak promo.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no measured run before this change; first run after it: 4% of the weekly limit per bench pass (W32 2026)
    • Claude Haiku 4.5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)
    • Claude Opus 5 on Claude Max 20x: no measured run before this change; first run after it: 1.5% of the weekly limit per bench pass (W32 2026)
    • Claude Sonnet 5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)

    Source: anthropics/claude-code #41930: Widespread abnormal usage limit drain since March 23, 2026

  3. AnthropicPromo

    Two weeks of doubled usage outside peak hours

    From March 13 Anthropic doubled limits outside weekday 8am to 2pm ET and all weekend, for Free, Pro, Max and Team. The bonus usage did not count toward weekly limits. The promo ended on March 28.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no measured run before this change; first run after it: 4% of the weekly limit per bench pass (W32 2026)
    • Claude Haiku 4.5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)
    • Claude Opus 5 on Claude Max 20x: no measured run before this change; first run after it: 1.5% of the weekly limit per bench pass (W32 2026)
    • Claude Sonnet 5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)

    Source: XDA: Claude just doubled every user's usage limits for two weeks, but only during off-peak hours

  4. GoogleLimit downunverified

    Antigravity AI Pro subscribers report multi-day lockouts

    Pro users hit the weekly cap early and stayed locked out for days. Forum threads and press coverage piled up through March; one source claims the weekly token allowance fell by about 97% between January and mid-2026, which is unconfirmed.

    Plans: Google AI Pro

    Measured
    • Gemini 3.1 Pro on Google AI Pro: no measured run before this change; first run after it: 16.8% of the weekly limit per bench pass (W32 2026)
    • Gemini 3.8 Flash on Google AI Pro: no measured run before this change; first run after it: 19.6% of the weekly limit per bench pass (W32 2026)

    Source: PiunikaWeb

January 2026

  1. AnthropicLimit downunverified

    Surprise limit complaints; Anthropic points to an expiring bonus

    Subscribers reported hitting limits sooner in the new year. Anthropic said a temporary bonus allowance had expired.

    Plans: Claude Pro, Claude Max 5x, Claude Max 20x

    Measured
    • Claude Fable 5.1 on Claude Max 20x: no measured run before this change; first run after it: 4% of the weekly limit per bench pass (W32 2026)
    • Claude Haiku 4.5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)
    • Claude Opus 5 on Claude Max 20x: no measured run before this change; first run after it: 1.5% of the weekly limit per bench pass (W32 2026)
    • Claude Sonnet 5 on Claude Max 20x: no measured run before this change; first run after it: 0.5% of the weekly limit per bench pass (W32 2026)

    Source: The Register