Five years ago, GPT-3 cost $0.06 per million output tokens and Anthropic did not exist as a consumer product. In 2026 the same dollar buys roughly 500x more reasoning, but only if you know which vendor is cheapest for which workload. This guide is the field-tested pricing map a working practitioner needs in September 2026: every major consumer subscription, every per-token API rate that matters, the code-editor tier, and the cost traps nobody mentions until the invoice arrives.
It is also worth saying the obvious: pricing pages change constantly. Every figure below was sourced from a vendor’s official pricing page between September 1 and September 21, 2026, or from Anthropic’s Claude Platform pricing docs, OpenAI’s API pricing page, or Google’s Gemini Developer pricing docs. Where a rate is announced to change after publication, the post says so in the text.
The 2026 AI subscription ladder at a glance
Five vendors now dominate the consumer AI subscription market. Three of them (Anthropic, OpenAI, Google) sell API access and subscriptions. Two of them (xAI, Perplexity) sell mostly subscriptions plus a developer API. The table below covers the subscription tier that a serious daily user actually pays for — the median bill.
| Vendor | Free tier | Standard paid | Power user tier | Annual discount |
|---|---|---|---|---|
| OpenAI (ChatGPT) | Free ($0, with ads in US) | Plus $20/mo | Pro $100 or $200/mo | None on Plus or Pro |
| Anthropic (Claude) | Free ($0) | Pro $17/mo ($200 annual) | Max from $100/mo (5x) | Yes, ~15% on annual |
| Google (Gemini) | Free ($0) | AI Pro $19.99/mo | AI Ultra from $99.99/mo (20x) | None published |
| xAI (Grok) | Free ($0, tight caps) | SuperGrok $30/mo | SuperGrok Heavy $300/mo | $300/yr on SuperGrok (~17%) |
| Perplexity | Free ($0) | Pro $20/mo ($200/yr) | Max $200/mo ($167/mo annual) | Yes, ~17% on annual |
The first thing to notice: four of the five vendors price their “default serious” tier between $17 and $30 a month. ChatGPT Plus ($20) has held at that exact number since 2023; Claude Pro dropped from $20 to $17 on annual billing; Gemini AI Pro sits at $19.99. Grok SuperGrok is the outlier at $30, and Perplexity Pro tracks the ChatGPT rate. The $20/month mark is the market reference point.
The second thing to notice: the “power user” tiers diverge wildly. Anthropic’s Max 5x starts at $100. OpenAI’s Pro tiers at $100 or $200. Google’s Ultra starts at $99.99 and tops out at $199.99. xAI’s SuperGrok Heavy is the most expensive consumer subscription in AI at $300/month. Perplexity Max is $200/month but with annual at $167. If you are paying more than $50/month, you are paying for something specific.
ChatGPT pricing in 2026: Free, Go, Plus, Pro, Business, Enterprise
OpenAI shipped a new $8/month tier called ChatGPT Go on January 16, 2026 — and announced in the same post that ads would start showing on Free and Go in the US. The middle of ChatGPT’s ladder is now genuinely crowded.
- Free ($0): Limited daily use, silent model downgrade after the cap (community-reported switch to GPT-5.3 mini). Since February 9, 2026 ads appear below responses in the US, labeled “Sponsored.” No opt-out beyond fewer daily free messages.
- Go ($8/mo): 10x more messages than Free, GPT-5.2 Instant access, longer memory. Ads also appear in the US. The right pick for a budget user who is OK with ads.
- Plus ($20/mo): Held at this price since launch. Full GPT-5.4 Thinking access, generous limits, Deep Research, Sora, Agent mode, custom GPTs, expanded voice. No ads.
- Pro ($100 or $200/mo): $100 buys 5x Plus usage; $200 buys 20x Plus usage with priority compute and the largest Deep Research budget. Power users only.
- Business ($20/seat/mo on annual, $25/seat/mo monthly): SAML SSO, SOC 2 Type 2, Slack/Google Drive connectors, business data excluded from training. Two-seat minimum.
- Enterprise (custom): Quote-based. SCIM, audit logs, custom data retention, the works.
The honest summary: Plus is still the right answer for almost everyone. Go makes sense if the $12/mo savings beats your tolerance for sponsored content. Pro is for people who genuinely hit the 80 messages/3-hour cap. Business is for teams of 4+ where admin controls and data-handling actually matter; at 4 seats you pay $80/mo annual for what looks like 4 Plus seats, but you get SOC 2 and SAML.
Three changes reset the free-vs-paid math this year: the launch of Go at $8 (lowered the floor), the February 9 launch of ads on Free + Go (made the floor less pleasant), and the April repricing of Business from $25 to $20 annual (made the team tier competitive with Plus). If you only pay for one AI subscription and you live in the US, Plus has the best cost-to-no-ads ratio.
Claude Pro, Max, Team: what $17 to $200 actually gets you
Anthropic restructured its consumer plans in 2026. Per the Anthropic pricing page, the lineup is now Free, Pro, Max (5x or 20x), and Team. The notable move: Pro went from $20/mo flat to $17/mo on annual billing ($200 paid up front), or $20/mo if billed monthly. That small annual discount is the first time Anthropic has offered a subscription discount to consumers.
- Free ($0): Full Claude access on web, desktop, and mobile, with up to 1M context on lower-tier models. No Claude Code, no Science, no Projects.
- Pro ($17/mo annual, $20 monthly): All Pro models including Claude Opus. Claude Code included. Research, Projects, Claude Design / Slides / Docs. Usage-credits system for the rare Fable tier.
- Max (from $100/mo): Two flavors — 5x or 20x the Pro usage. Choose 5x for $100/mo or 20x for $200/mo. Early access to advanced features, priority compute at peak, no usage credit deduction for Fable tier.
- Team ($20/seat/mo annual, $25/seat/mo monthly): Standard seat, more usage than Pro, central billing. Premium seat at $100/seat/mo annual (5x Pro usage) for power users inside an organization.
Max 5x at $100/mo has quietly become the new power-user default. It is cheaper than ChatGPT Pro at the same usage level, gives you Claude Opus (the strongest coding model in head-to-head benchmarks as of mid-2026), and includes Claude Code as part of the package. If you would otherwise pay $20 for ChatGPT Plus and $20 for Cursor, Max 5x at $100 might genuinely be more cost-effective at $100 total than the alternative $60 plus the lost Claude Code.
The annual discount on Pro is the pricing signal worth noting. Anthropic is the only frontier vendor offering any kind of annual deal at the consumer tier (Perplexity also offers annual, with the same ~17% effective discount). ChatGPT does not offer annual. Gemini does not offer annual on the AI plans (the Google One storage add-on is a separate product). Grok has the $300/year SuperGrok buy which is also ~17% off monthly.
Gemini prices drop 60% in 2026: Plus, Pro, Ultra
Google restructured Gemini pricing at I/O 2026 and the bill looks nothing like it did in 2025. Per the Google AI plans page, the consumer ladder has four tiers (Plus, Pro, Ultra, plus the bundled Workspace plans), and the headline cut is Google AI Ultra from $249.99/month down to $99.99/month — a 60% price cut on the top consumer tier. The price movement that Google billed as “the most aggressive subscription move in the AI era” is, in real terms, exactly that.
- Free ($0): Flash models only (Pro removed from free tier April 1, 2026 per Google pricing docs), 5-15 RPM, data used to improve Google products.
- Google AI Plus ($4.99/mo): New entry tier, cut from $7.99 on June 8, 2026. 400 GB storage, more access to Gemini, family sharing up to 5.
- Google AI Pro ($19.99/mo): Full Gemini 3.1 Pro access at 1M context, 5 TB storage, 1,000 AI credits/mo, YouTube Premium Lite, the best value in the lineup.
- Google AI Ultra (from $99.99/mo): 20 TB storage, ~5x Pro limits to start, 20x tier at $199.99/mo. Deep Think mode, Gemini Spark, Project Mariner, Gemini Agent in the US.
The Ultra cut is the headline; the strategic story is what sits underneath. Google removed Gemini Pro from the free tier, an explicit upmarket pivot. Gemini 3.1 Pro is the model Google is positioning as the agentic default, and the company is pricing Pro access so that it sits ~$0 below ChatGPT Plus while offering 1M context (versus ChatGPT Plus’s smaller practical context at the same tier). On paper, Pro is the most generous paid consumer tier in the market on a per-feature basis.
On the API side, several Gemini 3.x rates are temporarily discounted through December 31, 2026 according to Google’s Gemini Developer API pricing page: Gemini 3.8 Flash at $0.75/$3.75 per million input/output tokens (rising to $1.50/$7.50 from January 1, 2027), and Flash-Lite at $0.075/$0.375 (rising to $0.15/$0.75). Those are not the post-2027 rates; they are the current promotional rates. If you are planning to build on Gemini, lock in the architecture before January 1.
Per-token API pricing: the real cost story
Subscriptions have a fixed monthly cost and an unspecified usage ceiling. APIs have a variable cost per million tokens and no fixed cost. The crossover happens fast: if you would burn more than ~1,000 conversations a month on the Plus tier (or its equivalent), the API is usually the cheaper way to buy that much capacity. The per-token table below is the source of truth for that decision.
| Model | Input $/M | Output $/M | Cached Input $/M | Source |
|---|---|---|---|---|
| GPT-5.6 Luna | $0.20 | $1.20 | — | OpenAI API pricing |
| GPT-5.4 mini | $0.75 | $4.50 | $0.075 | OpenAI API pricing |
| GPT-5.4 | $2.50 | $15.00 | $0.25 | OpenAI API pricing |
| GPT-5.5 | $5.00 | $30.00 | $0.50 | OpenAI API pricing |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | Anthropic pricing |
| Claude Sonnet 5 | $2.00 | $10.00 | $0.20 | Anthropic pricing |
| Claude Opus 5 | $5.00 | $25.00 | $0.50 | Anthropic pricing |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 | Anthropic pricing |
| Gemini 3.1 Flash-Lite (current rate) | $0.25 | $1.50 | $0.075 | Gemini pricing |
| Gemini 3.5 Flash-Lite (current rate) | $0.30 | $2.50 | $0.075 | Gemini pricing |
| Gemini 3.8 Flash (current rate) | $0.75 | $3.75 | $0.075 | Gemini pricing |
| Gemini 3.1 Pro | $2.00 | $12.00 | — | Gemini pricing |
| Grok 4.3 | $1.25 | $2.50 | — | CloudZero tracking |
| Grok 4.5 | $2.00 | $6.00 | ~$0.50 | CloudZero tracking |
Three things stand out. First, on the absolute bottom of the table, GPT-5.6 Luna at $0.20/$1.20 is the cheapest frontier-tier model in the API market — full stop. If you have a workload that does not need frontier intelligence, Luna is the price floor.
Second, on the top of the table, all four frontier-model vendors now sit within 5% of each other at the high end: Claude Opus 5 at $5/$25, GPT-5.5 at $5/$30, Claude Fable 5.1 at $10/$50. The frontier is genuinely differentiated by capability, not price. If you need Sonnet 5’s price-to-quality ratio for production traffic, that is the seat to occupy at $2/$10 with batch bringing it down to $1/$5.
Third, cached input is where the real discounts are. Anthropic and OpenAI both offer 90% off the input rate for cached prompts, meaning repeated system prompts and large document context cost almost nothing. If you build any RAG or agentic system that sends the same tool schemas or system prompt repeatedly, prompt caching can reduce your monthly API bill by 50-70% without changing application logic.
Cursor, Copilot, and the code editor tier
The code editor AI tier is its own market, and it is priced very differently from general-purpose AI. Two vendors dominate: Cursor and GitHub Copilot. Both have restructured in 2026.
- Cursor Hobby ($0): No credit card, limited Agent requests, Composer access. For tinkerers.
- Cursor Individual ($20/mo): Extended Agent limits, frontier-model access, Grok Bot access, MCPs / skills / hooks, cloud agents, Bugbot on usage-based billing.
- Cursor Teams ($40/user/mo): Everything in Individual, plus centralized billing, team marketplace for skills/plugins, agentic code reviews with Bugbot, usage analytics, team privacy mode, SAML/OIDC.
- Cursor Enterprise (custom): Pooled usage, invoice/PO billing, SCIM, repository/model/MCP access controls, audit logs, priority support.
- GitHub Copilot Free ($0): Limited chat and agent usage; CLI and agent mode included; 2,000 inline suggestions per month.
- Copilot Pro ($15/mo with $10 in base AI credits): Unlimited inline suggestions, unlimited chat, code review, cloud agent, MCP integration.
- Copilot Pro+ ($70/mo with $39 in base credits): Adds delegation to third-party coding agents like Anthropic Claude and OpenAI Codex; $31 in additional flex credits.
- Copilot Max ($200/mo with $100 in base credits): Pooled org credits, larger flex allotment ($100 in additional credits), all environments.
Cursor’s Individual tier at $20 is directly comparable to ChatGPT Plus at $20: similar monthly spend, broad model access, strong editor integration. Cursor pulls ahead for users who want Claude Code as the editor’s primary completion model rather than a sidebar. GitHub Copilot Pro at $15 is the cheapest paid tier in this guide that gives you unlimited inline suggestions in the IDE, which is what most developers actually pay for. For a side-by-side feature and price comparison of all three code-editor AI tiers, see our Claude Code vs Cursor vs Copilot: The Honest 2026 AI Coding Showdown.
The interesting frontier is Copilot Pro+ at $70: that is the tier at which you can delegate to Claude Code or OpenAI Codex through Copilot’s own UI, with pooled credits. For a senior engineer running two or three of these agents in parallel, $70/mo for Pro+ plus a Max 5x Claude subscription at $100/mo is the canonical “I want the agentic web to work” combination. It is also $170/mo — not cheap.
Perplexity and the search tier
Perplexity is the fifth major consumer subscription tier, and its pricing model is the cleanest to explain. Per the Perplexity Pro page and Max page, there are four relevant plans.
- Free ($0): Unlimited standard searches with citations, no daily cap. ~5 Pro Searches per day. Sonar model only. No file uploads beyond ~3/day.
- Pro ($20/mo or $200/yr): ~300 Pro Searches per day, Deep Research, multi-model picker (GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.7 Flash, Kimi K3, GLM 5.3), 4,000 bonus credits, Sonar API access included.
- Max ($200/mo or $167/mo annual): Unlimited Pro Search + priority access. Model Council (3 models simultaneously). Perplexity Computer (research, design, code, ship workflows). 10,000 monthly credits.
- Enterprise Pro / Max: $34/seat/mo or $271/seat/mo annual. SSO, SOC 2, SCIM/audit logs gated by signed contract from August 14, 2026.
Pro at $20 is genuinely the best value in research-tier AI. The 300 daily Pro Searches exceed what any individual user will realistically burn. Max at $200 unlocks the Model Council feature that runs three frontier models in parallel and synthesizes a comparison — for high-stakes research tasks this is qualitatively different from any single-model query, and the $200 is rational if the work is paying for itself.
The Sonar API is the developer face of Perplexity. Per the UsagePricing calculator as of August 26, 2026, Sonar input runs about $1/1M tokens, Sonar Pro $3/1M. That is the ceiling-on-real-cost for an OpenAI-API-compatible search-augmented endpoint.
Five pricing traps that cost real money in 2026
Headline prices are not the prices you pay. Five failure modes have caused surprise invoices across the AI developer community this year.
Trap 1: Long-context premiums
Every flagship vendor has a “past this many tokens, double your rate” cliff. Per OpenAI’s pricing page, GPT-5.5 prompts over 272,000 input tokens are billed at 2x input and 1.5x output for the entire session, even under the Batch and Flex pricing tiers. Per Anthropic’s model docs, Claude Sonnet 4.5 prompts over 200,000 tokens move to premium long-context rates ($6 input / $22.50 output) for older Sonnet models. Per Google’s Gemini pricing page, Gemini 3.1 Pro more than doubles input cost past 200K tokens. If your context window routinely passes these cliffs, your monthly bill will be 1.5-2x your mental model of “input rate x tokens.”
Trap 2: Thinking tokens are billed as output
Reasoning models like Claude Opus with extended thinking, OpenAI’s o-series, and Gemini’s Deep Think mode silently inflate their bills. Claude explicitly documents this on its pricing page: “Extended thinking is included in output token cost.” On Claude Opus 4.8 that is $25 per million output tokens. On a workload where 60-80% of the billed output is hidden in the chain-of-thought, the effective per-reasoning-call cost can be 2-3x what the headline output rate suggests.
Trap 3: Batch API is non-interactive
OpenAI’s Batch API is 50% off both input and output. Anthropic’s batch is the same 50% discount. Google’s is the same 50% discount. The catch: Batch processing takes up to 24 hours to complete. If you are running a real-time product and someone hands you a workload expecting interactive latency, batching it 50% of the time produces a noticeable user-facing delay. The savings are real; the architectural fit has to be too.
Trap 4: Prompt cache writes cost more than you think
All three vendors now offer prompt caching with 90% off repeated input. The hidden half of the equation is the cache write: Anthropic’s pricing docs show 5-minute cache writes at 1.25x base input, and 1-hour cache writes at 2x base input. For Sonnet 5 at $2/M input, that is $2.50 and $4 per million cached-write tokens. If your architecture writes new caches aggressively, the write cost can swallow the read savings.
Trap 5: Ad-supported tiers are not free
ChatGPT Go at $8/mo looks like a third the price of Plus, but as TechCrunch reported on the rollout of ads in the free and Go tiers in February 2026, Go includes sponsored responses in the US. Anthropic responded with Super Bowl ads mocking the idea. For some users the ads are fine; for others they are a reason to stay at Plus. Same logic for any AI subscription that costs less than $20 — assume there is a hidden monetization layer above the sticker price, and decide whether you are OK with it.
How to pick what to pay for in 2026
A practical decision framework, in three steps.
- Identify your single daily-driver product. If you live in ChatGPT, ChatGPT Plus. In Claude, Claude Pro. In Cursor, Cursor Individual. In Copilot, Copilot Pro. Pay for the one you actually open every day — the one that fits your existing workflow. That single subscription handles 80% of your AI use.
- Add one API key for bursty workloads. The moment your daily-driver hits a usage cap (Plus on heavy research days, Pro on agent builds) the API is the cheaper escape hatch. Anthropic, OpenAI, and Google all offer pay-as-you-go with no commitment. A $20 spend on the API can clear a week of Plus-cap overflow.
- Audit quarterly. Pricing changes constantly. ChatGPT Go did not exist in 2025. Gemini Ultra cost $249.99 six months ago. Claude Pro dropped to $17 annual in 2026. Run a 15-minute pricing audit every quarter: check the official pricing pages, recompute your last month’s actual usage at the current rates, and switch if the math moved.
Self-hosted open-source models (Llama, Mistral, DeepSeek, Qwen) are not in this guide because the cost structure is fundamentally different: you pay electricity, GPU time, and engineering, not per-token rates. They are the right answer for workloads with steady high-volume traffic and a team that can run them. That is a different article.
Frequently asked questions
Is ChatGPT Plus still $20 in 2026?
Yes. Per OpenAI’s January 16, 2026 announcement, Plus has held at $20/mo since its 2023 launch. The new lower tier is ChatGPT Go at $8/mo, but that tier shows ads in the US.
What is the cheapest flagship-model API in 2026?
GPT-5.6 Luna at $0.20/$1.20 per million tokens, or Grok 4.3 at $1.25/$2.50. For frontier-grade reasoning at the lowest cost, Claude Haiku 4.5 at $1/$5 per Anthropic’s pricing docs is the canonical value pick.
Which AI subscription should I pay for if I only use one?
For general-purpose daily use, ChatGPT Plus ($20) or Claude Pro ($17 annual) is the right answer. For coding, Cursor $20/mo or GitHub Copilot Pro $15/mo is the right answer. For research-heavy work, Perplexity Pro $20/mo. Heavy power users can step up to Pro / Max / Ultra tiers at $100-$300/mo, but only if you actually hit the cheaper tier’s limits.
Are ads in ChatGPT real?
Yes. OpenAI began testing ads on Free and Go tiers in the US in early February 2026 — confirmed in TechCrunch’s coverage and OpenAI’s own advertising-approach post. Plus, Pro, Business, and Enterprise are ad-free.
Did Google cut Gemini prices in 2026?
Yes. Per the Google AI plans page, Google AI Ultra dropped from $249.99/mo to $99.99/mo at I/O 2026. Several Gemini 3.x API rates are also discounted through December 31, 2026 — including Gemini 3.8 Flash at $0.75/$3.75 per million tokens, scheduled to rise to $1.50/$7.50 from January 1, 2027.
What is the hidden cost most people miss?
Long-context premiums. OpenAI’s API pricing page documents that GPT-5.5 prompts over 272,000 input tokens are billed at 2x input and 1.5x output for the entire session. Anthropic’s Sonnet 4.5+ models bill 1.5x past 200K tokens on some plans. Google’s Gemini 3.1 Pro doubles input cost past 200K. These premiums can double a bill without warning.
Related reading
- Fine-Tuning vs RAG vs Prompt Engineering: When to Use Each AI Technique — practical decision framework for getting the most out of your API spend.
- LLM Context Windows: What 1M Tokens Actually Buys You — the post-training tax for long context, with the premiums explained.
- Claude Opus vs GPT vs Gemini: I Benchmarked All Three — which flagship gives the most intelligence per dollar.
- AI Safety Myths vs Reality: What the Field Actually Knows — what to watch for as vendors jockey on price.
This guide was sourced from each vendor’s official pricing page between September 1 and September 21, 2026. Pricing pages change frequently; for the most current figures, follow the links in each section. Last updated: 2026-09-21.