LLM Usage Limits 2026: ChatGPT vs. Claude vs. Gemini (Full Comparison)

If you’ve been using AI tools for a while, you already know the frustration: you find a workflow that clicks, you rely on it, and then one Tuesday morning something changes. A limit tightens. A plan gets restructured. A model you counted on gets sunsetted.

June took that pattern and turned the dial to eleven. A frontier model launched, got suspended by the US government, and came back with new limits — all inside three weeks. If you only skim one month of this guide all year, make it this one.

What you’ll find here is a current, practical breakdown of usage limits, pricing, models, and features across ChatGPT, Claude, Gemini, Grok, and Perplexity — as of July 2026. It’s not a “best AI” ranking. There’s no winner. Different platforms are better for different things, and your workflow probably looks nothing like the next person’s.

The goal is simple: enough accurate information to pick the right tool (or combination of tools) for what you actually do.

What’s Changed Since Late May 2026

June was the strangest month in AI subscriptions since this guide started. Here’s the short version.

Claude Fable 5 launched, vanished, and came back — with strings attached. Anthropic released Fable 5 on June 9 as its first “Mythos-class” model, a new tier above Opus with a 1M token context window. Three days later, a US government export-control directive forced Anthropic to suspend it entirely. On July 1 it returned globally with new safety classifiers — but Pro, Max, Team, and select Enterprise subscribers only get it for up to 50% of their weekly usage limits through July 7. After that, Fable 5 moves to usage credits billed at API-equivalent rates. Full story in the Claude section, because this one matters if you pay for Claude at any tier.

Claude Sonnet 5 is the new default for Free and Pro. Launched June 30 with near-Opus 4.8 agentic performance at a much lower cost. If you’re on Claude’s free tier or Pro, the model answering your everyday prompts just got a real upgrade — quietly, while everyone was watching the Fable drama.

Anthropic’s programmatic credit split is now live. The June 15 change we flagged last month happened on schedule. Agent SDK, claude -p, GitHub Actions, and third-party agent tools now draw from a separate monthly credit pool ($20 Pro / $100 Max 5x / $200 Max 20x) instead of your subscription quota. You had to claim the credit via email — if you run automations and haven’t, do it in your Claude account settings now.

OpenAI retired two model families. GPT-5.2 (Instant, Thinking, and Pro) left ChatGPT on June 12, and GPT-4.5 followed on June 26. Existing conversations auto-continue on GPT-5.5. o3 is next — it retires August 26. If a custom GPT or saved workflow still references any of these, it’s already broken or about to be.

Codex Remote went GA on every ChatGPT plan. You can now start or steer coding work on a connected Mac or Windows machine from the ChatGPT mobile app — review progress, approve actions, all from your phone. This includes Free and Go.

ChatGPT memory got real controls. You can now delete individual memories from the memory summary page or turn memory off entirely. Web is live; mobile is rolling out. Given how much personalization ChatGPT now does, this was overdue.

GPT-5.6 exists, but you can’t have it yet. OpenAI announced a limited preview of GPT-5.6 (three variants: Sol, Terra, Luna) on June 26 — available to roughly 20 vetted organizations via API and Codex, following a June 2 executive order that set up a government review process for new frontier models. General release is promised “in the coming weeks.” Nothing changes for ChatGPT subscribers today, but this is the same regulatory machinery that took Fable 5 offline, and it’s now shaping release calendars across the industry.

Gemini’s usage limits changed shape. Google’s shift from fixed daily prompt caps to compute-based limits is now rolling out broadly: your allowance refreshes every five hours until you hit a weekly ceiling, and complex prompts (video, coding, long chats) burn through it faster than simple ones. AI Pro and Ultra subscribers can buy pay-as-you-go top-up credits.

Gemini 3.5 Pro missed June. Google said “give us until next month” at I/O. The month came and went — 3.5 Pro is now expected in July, still in limited enterprise preview. Don’t plan around it yet.

Perplexity Pro isn’t “unlimited” anymore. Pricing-page snapshots now show metered allowances — roughly 200 Pro queries per week and 20 Deep Research runs per month on Pro, with the free tier at around 3 Pro Searches per day. Details in the Perplexity section.

Quick Comparison: All Five Platforms at a Glance

Usage limits on free and paid tiers aren’t always publicly disclosed and vary by demand, region, and account history. Data verified July 3, 2026.

Usage limits on free and paid tiers aren’t always publicly disclosed and vary by demand, region, and account history. Data verified July 3, 2026.

July 2026 pricing note: The $100 power tier is now a genuine three-way fight — ChatGPT Pro $100, Claude Max 5x, and Google AI Ultra at $99.99. But the bigger structural shift is what sits on top of every subscription: metered credits. Claude’s Fable 5 moves to usage credits after July 7. Anthropic’s programmatic pool went live June 15. Gemini sells top-up AI credits. ChatGPT sells Codex credits. The flat monthly fee increasingly buys you the floor, not the ceiling.

ChatGPT (OpenAI)

OpenAI spent June cleaning house. Two model families retired, memory controls shipped, Codex Remote went GA everywhere, and a next-generation model got announced that almost nobody can touch yet. The seven-tier plan structure is unchanged, which after the last few months counts as stability.

The Plans

Free ($0) gets you GPT-5.5 Instant access, basic file and image uploads, web browsing, and the ability to use (not create) custom GPTs. It’s limited, it degrades to a lighter model when you hit the cap, and — in the US, Australia, New Zealand, and Canada — it comes with ads. Your conversations may be used to train OpenAI’s models by default; opt out in Settings → Data Controls. New this month: Codex Remote works even on Free, so you can steer coding tasks on your own machine from the mobile app.

Go ($8/month) is the budget tier. More message volume than free, same ad situation. What it doesn’t include: advanced reasoning models, Sora, Deep Research, or Agent Mode. If you use ChatGPT for professional work, skip Go and go straight to Plus. The $12 difference buys you a completely different product.

Plus ($20/month) is where ChatGPT becomes a serious work tool. GPT-5.5 Thinking, 1M token context, Deep Research (10 runs/month), Sora video, Codex, Agent Mode, memory sources, and personalization from past chats and connected Gmail. The price hasn’t moved in three years while the product has expanded substantially. For most individuals doing real work, this is still the right tier.

Pro $100 ($100/month) gives you the same model suite as Pro $200 — including GPT-5.5 Pro — at 5x Plus usage instead of 20x. The practical difference is ceiling, not capability. If you’ve been bumping into Plus limits but nowhere near what the $200 plan is built for, this is your tier. Note that some models carry separate usage allowances on Pro tiers, and the $100 tier’s allowances are lower than the $200 tier’s — when you hit one, that model goes temporarily unavailable until the displayed reset time.

Pro $200 ($200/month) remains the top individual tier: 20x Plus usage, 1M token context in ChatGPT, and the most headroom for long reasoning and agentic sessions. US Pro subscribers also keep the personal finance preview (bank and brokerage connections via Plaid).

Business ($20/seat/month annual, $25 monthly) includes SAML SSO, an admin console, conversations excluded from training by default, and — new in June — plugin management in Workspace settings plus usage analytics and spend controls for admins.

Enterprise (custom pricing) adds SCIM, SLAs, compliance certifications, custom data retention, and an updated model picker that simplifies choosing speed vs. reasoning effort.

Models Right Now

GPT-5.5 is the flagship across all paid tiers, with GPT-5.5 Pro (higher-effort reasoning) on Pro and Business tiers and GPT-5.5 Instant as the default for everyone.

Retired in June: GPT-5.2 Instant, Thinking, and Pro left ChatGPT on June 12. GPT-4.5 left on June 26, including for custom GPTs. Conversations that used them continue automatically on GPT-5.5. o3 retires August 26 — the last call for anything still built on it. OpenAI’s stated policy: models generally stick around for 90 days after a successor ships. Plan accordingly.

GPT-5.6 (Sol, Terra, Luna) was announced June 26 as a limited preview — about 20 vetted organizations, API and Codex only, after OpenAI ran its release plans past the US government under the new executive-order framework. Sol targets the hardest problems, Terra high-volume business tasks, Luna fast/cheap everyday work. OpenAI says broader availability is coming soon and has publicly pushed back on government-gated access becoming the norm. Until it lands in ChatGPT, this changes nothing about which plan you should buy.

On Usage Limits

OpenAI is still notably vague about specific message caps. What’s consistently observed: Plus users have generous limits for most workflows, but complex multi-step reasoning or large-context tasks eat through them faster. There’s still no official dashboard showing remaining messages — you find out when you hit the wall. If you’re regularly hitting it, the $100 Pro tier is likely the answer before committing to $200.

Claude (Anthropic)

Claude had the most eventful month of any platform on this list — possibly the most eventful month any AI subscription has ever had. There are three separate stories here, and they affect different users differently. Take them in order.

Story one: the Fable 5 rollercoaster

On June 9, Anthropic launched Claude Fable 5 — the first model in a new “Mythos-class” tier that sits above Opus. 1M token context, always-on adaptive thinking, and benchmark results that led the field. It was included in Pro, Max, Team, and seat-based Enterprise plans, with an included window originally running through June 22.

Then it got weird. On June 12, a US government export-control directive forced Anthropic to suspend Fable 5 (and its restricted sibling, Mythos 5) for all customers. The trigger: Amazon researchers had found a way to prompt the model into demonstrating a software vulnerability exploit. Anthropic’s own testing found that plenty of other models — including GPT-5.5 and its own older Opus models — could produce the same demonstration, but the directive stood while the review played out. Every other Claude model kept working normally.

On July 1, Fable 5 came back globally with a new classifier that Anthropic says blocks the jailbreak in over 99% of attempts. Here’s the part that stings if you’re a subscriber: Pro, Max, Team, and select Enterprise plans get Fable 5 for up to 50% of weekly usage limits only through July 7. After that, it moves to usage credits billed at rates equivalent to the API’s $10/$50 per million tokens — double Opus 4.8, and roughly five times Sonnet 5’s intro rate. Subscribers who expected the original two-week included window are understandably annoyed; they got about three days before the suspension and now a week at half usage.

The silver lining: Anthropic says this is a capacity problem, not a permanent paywall, and that it intends to restore Fable 5 as a standard part of subscriptions “when sufficient capacity allows.” No date attached. If Fable 5 matters to your work, frontload heavy jobs before July 7, then route day-to-day work to Opus 4.8 or Sonnet 5 and save credits for the tasks that genuinely need frontier capability.

Story two: Sonnet 5 quietly raised the floor

While everyone watched Fable, Anthropic shipped Claude Sonnet 5 on June 30 — and made it the default model for Free and Pro plans. It plans, uses tools like browsers and terminals, and runs autonomously at a level that recently required much bigger models. On some knowledge-work benchmarks it edges out Opus 4.8. API pricing is $2/$10 per million tokens through August 31, then $3/$15.

For most Claude users, this matters more than Fable 5. Your everyday model just got meaningfully better at no extra cost.

Story three: the programmatic split is live

The June 15 billing change happened on schedule. Claude Agent SDK, claude -p, GitHub Actions workflows, and third-party agent tools no longer draw from your subscription quota. They use a separate monthly credit pool: $20 for Pro, $100 for Max 5x, $200 for Max 20x, metered at API list prices with no rollover. Credits had to be claimed once through your account settings — claim emails went out the week of June 8, and the credit doesn’t activate by itself.

Interactive use is unaffected: Claude.ai chat, Claude Code run directly in your terminal, and Claude Cowork all still draw from your normal subscription limits.

The Plans

Free ($0) now runs on Claude Sonnet 5 with daily limits that vary by demand. Memory from chat history is included. Limits are real — if you’re doing substantial work, you’ll feel them.

Pro ($20/month, $17/month annual) opens up the full tool suite: Claude Code, file creation and code execution, unlimited projects, Google Workspace integration, remote MCP connectors, Claude Design, Sonnet 5 as the default, and Opus 4.8 for harder problems. Fable 5 access through July 7 (50% of weekly limits), then via credits. About 5x more usage than free.

Max ($100/month or $200/month) is built for people who routinely bump into Pro’s caps — 5x or 20x usage, the doubled 5-hour session rate limit from May, persistent memory across conversations, Claude Code Auto mode, and early feature access. Same Fable 5 terms as Pro.

Team ($25/seat/month annual, $30 monthly) adds collaboration, shared projects, and admin controls. Premium seats at $150/month add the full Claude Code environment. Same Fable 5 window as Pro/Max.

Enterprise (custom) — worth knowing: premium Enterprise seats share the July 7 Fable grace window, but standard Enterprise seats need usage credits enabled from day one to run Fable 5 at all.

Models Right Now

Haiku 4.5 — fastest and cheapest ($1/$5 per MTok) for high-volume tasks. Sonnet 5 — the new default for Free and Pro. Near-Opus agentic performance. $2/$10 intro through Aug 31, then $3/$15. Opus 4.8 — the Opus-tier flagship ($5/$25), still the accuracy pick for the hardest agentic work. Fable 5 — the Mythos-class ceiling ($10/$50), 1M context, metered via credits after July 7.

Also new June 30: Claude Science, a platform aimed at research workflows — literature synthesis, experimental planning, lab-tool integration. Early days, but if you work in research it’s worth a look.

Claude still doesn’t generate images. Pair it with something else for visuals, or use Claude Design for the visual work it does handle.

Gemini (Google)

Google’s month was quieter than Anthropic’s, but one change affects every Gemini user: the way limits are counted has fundamentally changed. Meanwhile the model everyone’s waiting for slipped its deadline.

The compute-based limits shift

Gemini has moved from fixed daily prompt caps to compute-used limits. Your allowance now factors in prompt complexity, the features you use, and chat length — a simple text question costs far less of your allowance than a video generation or a long coding session. The allowance refreshes every five hours until you hit a weekly ceiling. Hit the cap on the biggest models and Gemini shifts you to its smaller, faster ones automatically. AI Pro and Ultra subscribers can buy pay-as-you-go top-up AI credits to keep going.

Honest take: this is probably fairer in aggregate, and it’s definitely harder to predict. You can no longer count prompts to know where you stand. If your usage is heavy and spiky — big research days, video generation bursts — expect to hit the weekly ceiling in ways daily caps never showed you.

The Plans

Free gives you Gemini 3 Flash, restricted Deep Research (5 reports/month), Gemini Live voice, Canvas, Gems, 100 monthly AI credits, and NotebookLM. Still a genuinely useful free tier — but Pro-class models remain paid-only since April 1, and the new compute-based limits apply here too.

Google AI Plus ($7.99/month) expands 3.1 Pro access beyond free, adds limited Veo 3.1 Lite video, Nano Banana Pro images, Gemini in Gmail, and 200 GB storage. The budget bridge tier.

Google AI Pro ($19.99/month, first month free) is the full package for most users: Gemini 3.1 Pro, expanded Deep Research, 1M token context, 1,000 monthly AI credits, Veo 3.1 video, Workspace integration across Gmail/Docs/Sheets, and higher limits in Code Assist and the Jules coding agent.

Google AI Ultra ($99.99/month or $200/month) — the I/O price cut stands. $99.99 gets 5x Pro usage, Deep Think, Veo 3.1, 10,000+ monthly AI credits, and Gemini Spark (the 24/7 agent, US beta). The $200 tier goes to 20x usage, 20 TB storage, and YouTube Premium.

Models Right Now

Gemini 3.1 Pro remains the consumer flagship. Gemini 3 Flash is the free-tier default, with Gemini 3.5 Flash (launched May 19) increasingly handling everyday work — it beats 3.1 Pro on coding and agentic benchmarks at lower cost.

Gemini 3.5 Pro is late. Announced at I/O with a “give us until next month” promise, it missed June entirely and is now expected in July — still in limited Vertex enterprise preview, with reports pointing to quality refinements around coding and token efficiency. The 2M token context window and Deep Think mode are worth waiting for, but treat the July date as a target, not a commitment. Google has missed two delivery windows this year.

Housekeeping: Gemini 2.0 Flash and 2.0 Flash-Lite shut down June 1 as scheduled. Veo 3 and Veo 2 API models followed June 30. If you built on any of these, you’ve already migrated or broken.

What’s Worth Knowing

Gemini’s real advantage is still the integration, not the model. Asking your Gmail inbox for an AI overview, working inside Docs and Sheets, pushing Deep Research outputs to Drive — no other platform matches that. If your day already runs on Google’s tools, the compounding value is real.

Grok (xAI)

The quietest month on the list. No pricing changes, no new consumer tiers, no retirements. The Grok 4.3 staged rollout that began in May is still working through SuperGrok and X Premium+ accounts, which means two subscribers on the same plan can still hit different model versions.

The Plans (unchanged since May)

Free — limited access, widely estimated around 10 requests per two hours. SuperGrok Lite ($10/month) — entry paid tier with basic image/video generation and one AI agent. SuperGrok ($30/month) — full 4.3 access (staged), DeepSearch, Big Brain Mode, priority routing. X Premium ($8) / Premium+ ($40) — Grok bundled with X platform features. Grok Business ($30/seat) — team controls, no training on your data. SuperGrok Heavy ($300/month) — Grok 4 Heavy multi-agent system, confirmed full 4.3 access, maximum limits, Custom Skills.

Models Right Now

Grok 4.3 is the default (1M context, native video input). The Grok 4.20 Multi-Agent variant still holds the largest context window anywhere: 2M tokens. Grok Build 0.1 handles automated coding pipelines at lower per-token cost. Grok 4.5 is reportedly in private beta on xAI’s new V9 base — no confirmed pricing or date, so treat it as rumor-adjacent for now.

What’s Worth Knowing

The real-time X data access remains genuinely unique, and the 2M context window is a real differentiator for whole-codebase or long-archive work. The $30/month price is still harder to justify against the $20 competition unless one of those two things is specifically what you need. And there’s still no persistent memory at any tier — the biggest gap versus every other platform here.

Perplexity

Perplexity’s product didn’t change much in June, but its limits language did — and it’s worth a correction from what we reported previously.

The “unlimited” era is over

Earlier this year, Pro was marketed around unlimited Pro Search. Current pricing-page snapshots tell a different story: Pro now lists roughly 200 Pro queries per week, 20 Deep Research runs per month, and 40 Comet Agent queries per month. The free tier shows about 3 Pro Searches per day, down from the long-standing 5. Sources conflict on exactly when this shifted — Perplexity didn’t announce it loudly — so check the usage panel in your own account for your actual numbers. The pattern matches the industry: metered allowances replacing “unlimited” promises everywhere.

The Plans

Free — unlimited standard searches with citations, ~3 Pro Searches/day, 1 Research query/month. Still genuinely usable for casual research. Pro ($20/month, $200/year) — the metered-but-generous tier: ~200 Pro queries/week, 20 Deep Research/month, model switching between GPT-5.4/5.5, Claude Opus 4.8, and Gemini 3.1 Pro, file uploads, and Perplexity Computer access with a one-time credit allocation. Max ($200/month) — unlimited Labs, Perplexity Computer with 10,000 monthly credits, Model Council multi-model validation, priority access. Education Pro ($10/month) for verified students. Enterprise Pro ($40/seat) / Enterprise Max ($325/seat) for teams.

What’s Worth Knowing

Comet stays free across iOS, Android, Windows, and Mac, with Comet Plus ($5/month publisher content) included in Pro and Max. Model switching remains the sleeper feature — comparing how three frontier models handle the same research query without paying three subscriptions. And it’s still not a general-purpose assistant: use it for research and verification, pair it with Claude or ChatGPT for creation.

One watch item: the EU AI Act’s general-purpose AI obligations take effect August 2. Perplexity hasn’t published a compliance statement, and EU users may see feature adjustments. We’ll cover it next month if anything moves.

Understanding New Usage Paradigms

Every platform manages usage differently, and June added two entirely new dynamics. The practical picture:

Rolling windows vs. compute budgets. ChatGPT uses rolling time windows per model. Claude uses a 5-hour rolling window plus weekly limits. Gemini now uses compute-based allowances — 5-hour refresh, weekly ceiling, complexity-weighted. Perplexity meters weekly Pro queries. Grok publishes almost nothing. The common thread: simple prompt-counting no longer tells you where you stand on most platforms.

The subscription + credits hybrid is the new normal. This is June’s biggest structural story. Flat subscriptions increasingly cover baseline usage, with metered credits layered on top for premium capability: Fable 5 usage credits and the programmatic pool on Claude, Codex credits on ChatGPT, top-up AI credits on Gemini, Computer credits on Perplexity. The “all-you-can-eat” AI subscription is quietly dying in the agent era, because an autonomous agent can burn compute thousands of times faster than a human typing. Budget accordingly: your monthly fee is the floor, and heavy months will cost more.

Governments are now a release variable. The June 2 US executive order created a review process for frontier models — and within a month it had suspended one shipping model (Fable 5, June 12–July 1) and gated another’s launch (GPT-5.6’s 20-organization preview). Whatever you think of the policy, the practical takeaway for users is new: a model you rely on can now disappear for reasons that have nothing to do with the company that makes it. Redundancy across platforms just became a resilience strategy, not only a quality one.

“Unlimited” means different things. No platform offers truly unlimited usage on any tier — and this month Perplexity’s Pro tier formally joined the metered club. “Unlimited” in practice means “high enough that most users won’t hit it,” subject to abuse prevention and load.

Context window vs. conversation limit. A large context window tells you how much fits in a single conversation. It says nothing about how many conversations you get per day. These are different constraints, and confusing them is still the most common mistake we see.

Peak-hour throttling is real. Claude, Gemini, and ChatGPT all degrade during high-demand windows. This is platform-wide load, not your personal cap. If your heaviest work happens during US morning hours, build in flexibility.

For an in-depth guide on premium-tier economics ($100–$300 plans and who actually needs them), see our companion article: Premium Tier Deep Dive (link). For tactics on stretching your allowances, see: Optimizing Your AI Strategy Around Usage Limits (link).

The Bottom Line

Start with what you do most. Agentic coding → Claude (and Sonnet 5 just made the $20 tier better). Broad feature set and integrations → ChatGPT. Living inside Google’s tools → Gemini. Research and source verification → Perplexity. Real-time X data or massive context → Grok.

The $20/month tier is still the sweet spot for most people — and it quietly improved this month. Claude Pro now defaults to Sonnet 5. ChatGPT Plus keeps expanding. Gemini Pro’s compute-based limits are unpredictable but generous for typical use.

The $100 tier is where the real competition is. ChatGPT Pro $100, Claude Max 5x, and Google AI Ultra $99.99 are now direct rivals. Evaluate here before jumping to $200.

Budget for credits, not just subscriptions. If you use frontier models heavily or run automations, your real monthly cost is subscription plus credits. Watch your usage dashboards — all five platforms now have meters that matter.

Use more than one. June proved a new reason why: a government directive took the industry’s newest flagship offline for three weeks with no warning. Most serious AI users already maintain 2–3 subscriptions for capability reasons. Resilience is now a second reason.

For persona-based recommendations (casual, regular, power, developer, business), see our companion article: Strategic Planning for Different User Types (link).

The Evolution Continues

Staying current with this stuff is genuinely hard — that’s why this article exists and updates monthly. Three tips that keep working: check the platform’s official pricing page before subscribing (things shift mid-month now), watch your usage dashboard rather than trusting remembered limits, and treat any model you depend on as temporary — the average flagship now lasts about 90 days before a successor arrives.

Key Takeaways

• Claude Fable 5 launched June 9, was suspended by a US export-control directive June 12, and returned July 1 — included at 50% of weekly limits through July 7, then metered via usage credits
• Claude Sonnet 5 is the new default for Free and Pro — near-Opus performance at the everyday tier
• Anthropic’s programmatic credit pool went live June 15 — claim your credit if you run automations
• GPT-5.2 and GPT-4.5 are retired from ChatGPT; o3 goes August 26
• Codex Remote is GA on every ChatGPT plan, including Free
• ChatGPT memory can now be edited, deleted, or turned off entirely
• GPT-5.6 is in a ~20-organization preview under the new US government review framework — not in ChatGPT yet
• Gemini limits are now compute-based — 5-hour refresh, weekly cap, top-up credits for Pro/Ultra
• Gemini 3.5 Pro slipped from June to July
• Perplexity Pro is no longer “unlimited” — ~200 Pro queries/week, 20 Deep Research/month

Pricing and features verified July 3, 2026. This article is updated monthly. All figures are approximate — always verify current pricing on each platform’s official page before subscribing.

Official sources: OpenAI release notes · Anthropic news · Google AI plans · xAI docs · Perplexity pricing

Related reading:

• NotebookLM vs Gemini for Research
• Claude vs ChatGPT for Coding
• AI File Upload Guide: ChatGPT, Claude, Gemini & NotebookLM
• How to Export from Claude
• Are AI Browsers Safe?