📖 10 min read
Quick Verdict: For pure price-per-token in 2026, Gemini 2.5 Pro wins on input cost ($1.25/M vs Claude Sonnet 4.5’s $3/M and GPT-5’s $1.25/M but with worse output economics). Claude Sonnet 4.5 wins for production workloads thanks to prompt caching (90% off) and best-in-class output quality per dollar. GPT-5.5 wins for cutting-edge reasoning if cost is no object. Best overall value: Gemini 2.5 Pro for cheap experiments, Claude Sonnet 4.5 for production, GPT-5 mini for high-volume simple tasks. Rating: 8.5/10 — pricing is genuinely competitive now; the old “Claude is 5x GPT-4” gap is gone.
Updated: June 2026
Pricing for the big three frontier models has shifted three times in the last 14 months. OpenAI released GPT-5 in August 2025, then GPT-5.5 in April 2026. Anthropic shipped Claude Sonnet 4.5 in September 2025, then quietly bumped Sonnet to 4.6 and Opus to 4.7 in May 2026 without changing headline rates. Google kept Gemini 2.5 Pro on a stable $1.25/$10 price since launch in June 2025.
This article is the 3-way per-token comparison we wish existed when we built our last AI product. No marketing fluff, no affiliate-driven “GPT-5 is amazing” framing. Just the numbers, the workload math, and where each vendor’s pricing is the actual winner.
The Headline Numbers
All three vendors price on a per-million-token basis, split between input tokens (what you send) and output tokens (what the model generates). Output is always 4-10x more expensive than input across all three vendors — and that’s the single most important fact in this entire article.
Here’s the current headline pricing for the flagship model of each vendor as of June 2026:
- OpenAI GPT-5: $1.25 per million input tokens, $10.00 per million output tokens
- Anthropic Claude Sonnet 4.5/4.6: $3.00 per million input tokens, $15.00 per million output tokens
- Google Gemini 2.5 Pro: $1.25 per million input tokens, $10.00 per million output tokens (≤200K context); $2.50 input / $15.00 output above 200K
On raw input cost, GPT-5 and Gemini 2.5 Pro are tied as the cheapest frontier-tier option at $1.25/M input. Claude Sonnet is 2.4x more expensive on input. But input is rarely where you’ll spend the bulk of your bill. Output is where the action is.
On output cost: GPT-5 wins at $10/M, Gemini 2.5 Pro ties at $10/M (under 200K context), and Claude Sonnet is 50% more expensive at $15/M. If your workload is output-heavy — content generation, code synthesis, agentic loops — that 50% premium adds up fast.
Input vs Output: Why the Asymmetry Matters
The reason output costs more is mechanical: output tokens require autoregressive generation, which means each token depends on every previous one. Input tokens are processed in parallel batches. The vendor’s compute cost is genuinely higher for output, and the price reflects it.
The asymmetry shapes how you should think about cost. Here’s a rough breakdown of common workload shapes:
- Summarization (PDF, articles, transcripts): 90% input / 10% output. You’re shoving in 50 pages and getting back a paragraph. Input cost dominates.
- Long-form generation (blog posts, reports, code from spec): 10% input / 90% output. Short prompt, 2,000 tokens back. Output cost dominates by 5-10x.
- Classification / extraction (100 emails, 1,000 invoices): 70% input / 30% output. Big batch in, structured JSON out. Input dominates but output is non-trivial.
- Agentic / tool-calling workflows: Variable, but commonly 40% input / 60% output. Multi-turn conversations with function calls. Mixed.
For a content-generation app that bills $50/month to a customer, your API cost on Claude Sonnet 4.5 is going to be about 1.5x what it would be on GPT-5 or Gemini 2.5 Pro, assuming identical output volume. For a PDF summarizer with high input and tiny output, the three are within 10% of each other on a per-task basis.
The other asymmetry people miss: prompt caching. Anthropic offers up to 90% off on cached input tokens (writes cost 25% more, but cached reads are $0.30/M on Sonnet 4.5). OpenAI offers cached input at 50% off. Google offers explicit context caching at the same $1.25/M rate but only after a minimum token threshold. If you re-use the same system prompt across thousands of requests — most production apps do — the cached input rate is what you actually pay on 80%+ of your input tokens.
Side-by-Side Price Table
| Model | Input $/M | Output $/M | Batch 50% off? | Cache discount | Context window |
|---|---|---|---|---|---|
| GPT-5 mini | $0.25 | $2.00 | Yes | 50% off cached input | 128K |
| GPT-5 | $1.25 | $10.00 | Yes | 50% off cached input | 272K |
| GPT-5.5 (flagship) | $5.00 | $30.00 | Yes | 50% off cached input | 1.05M |
| Claude Haiku 4.5 | $1.00 | $5.00 | Yes (50%) | Up to 90% off cached | 200K |
| Claude Sonnet 4.5/4.6 | $3.00 | $15.00 | Yes (50%) | Up to 90% off cached | 200K (1M beta) |
| Claude Opus 4.7 | $5.00 | $25.00 | Yes (50%) | Up to 90% off cached | 200K (1M beta) |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | Yes (50%) | Context caching free tier | 1M |
| Gemini 2.5 Flash | $0.30 | $2.50 | Yes (50%) | Free caching (under thresholds) | 1M |
| Gemini 2.5 Pro | $1.25 (≤200K) $2.50 (>200K) |
$10.00 (≤200K) $15.00 (>200K) |
No batch discount on Pro | Explicit context cache | 1M (2M Pro Experimental) |
Three observations from this table:
1. The cheap tiers are shockingly cheap. Gemini 2.5 Flash-Lite at $0.10/M input and $0.40/M output is genuinely 10x cheaper than GPT-5 mini. If you’re running a classification pipeline that doesn’t need frontier reasoning, Flash-Lite is the cheapest credible model on the market.
2. Claude’s caching is the killer feature if you’re running at scale. Anthropic’s 90% off cached input means that for a chatbot with a fixed 5,000-token system prompt, your effective input cost on Sonnet 4.5 drops from $3.00/M to $0.30/M after the first request. No other vendor matches that.
3. OpenAI’s batch API is the killer feature if you’re running async jobs. 50% off on both input and output, with a 24-hour SLA. For nightly batch summarization, OpenAI is the cheapest path. Anthropic has batch too, but Gemini 2.5 Pro notably does not — Google prefers you use their async prediction endpoint at standard rates.
Real Workloads: Same Prompt, Real Cost
Pricing tables are abstract. Let’s run four real workloads through each platform and compute the actual bill. We’re assuming standard non-cached, non-batch rates and standard context (under 200K).
Workload 1: Summarize a 50-page PDF
Input: ~75,000 tokens (PDF text + system prompt). Output: ~800 tokens (summary).
– GPT-5: (75,000 × $1.25) + (800 × $10) = $0.094 + $0.008 = $0.10
– Claude Sonnet 4.5: (75,000 × $3.00) + (800 × $15) = $0.225 + $0.012 = $0.24
– Gemini 2.5 Pro: same math as GPT-5 = $0.10
Workload 2: Write a 2,000-word blog post from a brief
Input: ~2,000 tokens. Output: ~2,800 tokens (≈2,000 words).
– GPT-5: (2,000 × $1.25) + (2,800 × $10) = $0.0025 + $0.028 = $0.03
– Claude Sonnet 4.5: (2,000 × $3.00) + (2,800 × $15) = $0.006 + $0.042 = $0.05
– Gemini 2.5 Pro: $0.03
Workload 3: Extract structured data from 100 emails
Input: ~150,000 tokens (100 emails × ~1,500 tokens each + schema prompt). Output: ~20,000 tokens (JSON for each).
– GPT-5: (150,000 × $1.25) + (20,000 × $10) = $0.1875 + $0.20 = $0.39
– Claude Sonnet 4.5: (150,000 × $3.00) + (20,000 × $15) = $0.45 + $0.30 = $0.75
– Gemini 2.5 Pro: $0.39
Workload 4: Generate 5,000 lines of Python from a spec
Input: ~8,000 tokens. Output: ~25,000 tokens.
– GPT-5: (8,000 × $1.25) + (25,000 × $10) = $0.01 + $0.25 = $0.26
– Claude Sonnet 4.5: (8,000 × $3.00) + (25,000 × $15) = $0.024 + $0.375 = $0.40
– Gemini 2.5 Pro: $0.26
Summary across the four workloads:
– Total on GPT-5: $0.78
– Total on Claude Sonnet 4.5: $1.44 (1.85x more expensive)
– Total on Gemini 2.5 Pro: $0.78 (tied with GPT-5)
If you cache the system prompt on Claude — say the system prompt is 1,500 tokens reused across all four — the cached input drops Sonnet’s effective input rate to $0.30/M, and the total falls to $1.10 across all four workloads. Still more expensive than GPT-5, but the gap closes from 85% premium to 41% premium.
Where Each One Wins
OpenAI GPT-5 wins for:
- Cheapest raw output at the frontier tier ($10/M). Best for content generation at scale.
- Batch API: 50% off both input and output for async workloads.
- GPT-5 mini at $0.25/$2 is the cheapest credible model for high-volume classification, routing, and simple extraction.
- Long-context (272K standard, GPT-5.5 hits 1M+) at flat rates, no tiered pricing jumps.
- Ecosystem: most third-party tools, SDKs, and integrations target the OpenAI API spec first.
Anthropic Claude Sonnet 4.5 wins for:
- Production chatbots and customer-facing apps: prompt caching cuts effective input cost by up to 90% on repeated system prompts.
- Code generation and agentic workflows: Sonnet 4.5’s output quality per dollar is widely regarded as best-in-class for structured reasoning tasks.
- Tool calling and multi-step planning: the output pricing premium pays for itself if the model gets the call graph right on the first try (fewer retries).
- Long-context workloads via batch API: 50% off plus 1M-token context (beta) is competitive with Gemini.
- Compliance-sensitive workloads: Anthropic’s enterprise posture and data retention defaults are stricter than OpenAI’s.
Google Gemini 2.5 Pro wins for:
- Cheapest frontier-tier input at $1.25/M, tied with GPT-5.
- Largest stable context window at 1M-2M tokens for Pro Experimental.
- Multimodal native: video, audio, image input at no premium. Other vendors charge or limit.
- Free tier: Gemini 2.5 Pro is still free in AI Studio with rate limits — best free frontier model in 2026.
- Flash-Lite at $0.10/M input is unbeatable for high-volume, low-stakes classification.
Hidden Costs: Rate Limits, Burst, and Quota Traps
The published per-token price is what you pay when everything goes smoothly. In practice, three things make your real cost higher than the table suggests.
1. Rate limits force retries. OpenAI’s tier-1 GPT-5 account is capped at 30,000 TPM (tokens per minute). If your app bursts above that, you get 429s and need to retry. Retry logic adds latency and, if you’re using a paid retry strategy, doubles token consumption on failures. Claude’s tier-1 limits are tighter (~20,000 TPM for Sonnet on default tiers). Gemini’s limits are looser but tier-gated — 2 RPM on the free tier, scaling up only after billing is enabled.
2. Quota tiers lock you into higher prices. OpenAI’s $50/month minimum commitment gets you tier-3 limits; below that you’re rate-limited to a degree that makes production usage impractical. If you’re paying $1.25/M input on paper but burning through retries at 2x token count, your effective rate is $2.50/M. Anthropic’s tier structure is similar: $5 minimum deposit gets you in the door, but production-grade limits require monthly commitments that effectively anchor you to ~$100/month minimums.
3. Output token inflation from “thinking” models. GPT-5.5 and Claude Opus 4.7 are reasoning models — they generate internal chain-of-thought tokens before the user-visible output. Those thinking tokens are billed as output tokens at the full rate. A simple “summarize this” prompt can return 2,000 tokens of visible output and 8,000 tokens of hidden reasoning, all at $30/M. That’s 4x the cost you’d estimate from the visible output alone. If you’re paying for “smarter” models, factor in the hidden reasoning tax.
4. Regional markups. OpenAI added a 10% uplift in March 2026 for data-residency endpoints (EU, US-only). If your enterprise contract requires regional processing, your headline rate becomes $1.375/M on GPT-5, $3.30/M on Sonnet (Anthropic’s regional pricing varies but averages 5-15% uplift), and Google has no current regional surcharge as of June 2026.
When to Use the API vs the Subscription
Each vendor offers a consumer subscription that’s cheaper than the API for moderate users:
- ChatGPT Plus: $20/month, includes GPT-5 access with generous usage caps (roughly 200 messages/3 hours before throttling).
- Claude Pro: $20/month, includes Sonnet 4.5/4.6 with ~5x more usage than the free tier.
- Google AI Studio + Gemini app: Free for Gemini 2.5 Pro with rate limits; Gemini Advanced at $20/month with 2TB storage.
The breakeven math: at $20/month subscription, you’re “pre-paying” roughly 2 million output tokens at GPT-5’s rate ($20 / $10/M = 2M tokens) or 1.33 million tokens at Claude Sonnet’s rate ($20 / $15/M = 1.33M tokens). If your monthly output token usage exceeds that, the API is cheaper.
Practical breakeven:
- ChatGPT Plus → GPT-5 API: Around 150-200 messages/day of meaningful length, or any automation / API integration. Below that, the subscription is the better deal.
- Claude Pro → Sonnet API: Around 80-120 messages/day. Sonnet’s higher per-token rate lowers the breakeven point.
- Gemini free → Gemini API: Gemini’s free tier is generous — 2 RPM and 32K tokens/minute. Most casual users will never need to pay.
For business use — anything with automated workflows, scheduled jobs, or multi-tenant serving — the API is always cheaper than subscription once you cross ~100 daily active interactions. The subscription makes sense for individual power users who want the model in a polished UI; it does not make sense for any workflow where you’re piping API calls.
Final Verdict
The 2026 frontier-model pricing landscape is genuinely competitive. The “Claude is too expensive” narrative from 2024 is outdated — Claude Sonnet 4.5 at $3/$15 with 90% prompt caching is a strong value proposition for production workloads with stable system prompts.
By use case, the winner is:
- Cheapest output for content generation: GPT-5 and Gemini 2.5 Pro tied at $10/M. Both beat Claude Sonnet by 50%.
- Best value for production chat apps with caching: Claude Sonnet 4.5. The 90% cache discount makes its effective rate the lowest among the three once you factor in prompt reuse.
- Cheapest at extreme scale (millions of simple requests): Gemini 2.5 Flash-Lite at $0.10/$0.40 is in a class by itself.
- Best for multimodal (image, video, audio input): Gemini 2.5 Pro. Native multimodal with no input premium.
- Best for reasoning-heavy tasks where cost is secondary: GPT-5.5 or Claude Opus 4.7. Both ~$5/$25-$30. Pick based on your eval benchmarks, not price.
- Best free option: Gemini 2.5 Pro in AI Studio, hands down.
For most builders shipping in mid-2026: prototype on Gemini 2.5 Pro (free tier), validate on GPT-5 (cheapest output at frontier tier), ship on Claude Sonnet 4.5 (best caching economics for production). The three are now close enough that quality and latency benchmarks should drive vendor choice — price is the secondary filter it should always have been.