Why this page exists
Every frontier vendor cut API prices in 2025. OpenAI shipped GPT-5.5 at $5/$30, Anthropic shipped Opus 4.7 and Sonnet 4.6, Google shipped Gemini 3.1 Pro and killed Flash-Lite on June 1. The price spread between the cheapest and most expensive frontier model is now roughly 1,250x — and it changes every 3–6 weeks. Without a single source of truth, you end up overpaying by 3–10x on the same workload because you decided prices six months ago and forgot to revisit them.
This hub is the page I check on the 1st of every month. It tracks every rate card for OpenAI, Anthropic, and Google, plus the open-weight hosts (DeepSeek, Qwen, Llama) where it makes sense. Below is the current rate card for July 2026, what’s new this month, a vendor decision tree, and links to the per-vendor deep dives.
Update cadence: this page is rewritten on the 1st of every month. Per-vendor deep dives are updated in lockstep when rates change. Last updated: July 3, 2026. Next refresh: August 1, 2026.
The current rate card (July 2026)
Every rate below is per 1M tokens. Cached input is shown in parentheses where supported. Long-context rates (>200K) apply where noted.
OpenAI
| Model | Input | Cached input | Output | Notes |
|---|---|---|---|---|
gpt-5.5 |
$5.00 | $0.50 | $30.00 | Flagship; long-context $10/$45 |
gpt-5.5-pro |
$30.00 | — | $180.00 | Top-of-stack; long-context $60/$270 |
gpt-5.4 |
$2.50 | $0.25 | $15.00 | Default mid-tier; long-context $5/$22.50 |
gpt-5.4-mini |
$0.75 | $0.075 | $4.50 | High-volume classification |
gpt-5.4-nano |
$0.20 | $0.02 | $1.25 | Cheapest OpenAI; fine-tunable |
gpt-5 (legacy stable) |
$1.25 | $0.125 | $10.00 | Predictable; quality sufficient for extraction |
o3-deep-research |
$10.00 | $2.50 | $40.00 | Deep research agent |
gpt-image-1.5 medium |
$5.00 | $1.25 | $10.00 | ~$0.02–$0.06 per 1024×1024 image |
gpt-realtime-2 |
$32.00 | $0.40 | $64.00 | $0.034/min translate |
Anthropic
| Model | Input | Cached input | Output | Notes |
|---|---|---|---|---|
claude-opus-4-7 |
$5.00 | $0.50 | $25.00 | Flagship reasoning; vision included |
claude-opus-4-6 |
$5.00 | $0.50 | $25.00 | Stable Opus tier |
claude-sonnet-4-6 |
$3.00 ($2.00 until Aug 31) | $0.30 | $15.00 ($10.00 until Aug 31) | Default mid-tier; intro rate ends Aug 31, 2026 |
claude-sonnet-4-5 |
$3.00 | $0.30 | $15.00 | Legacy Sonnet |
claude-haiku-4-5 |
$1.00 | $0.10 | $5.00 | Cheapest current Claude |
| Batch API discount | ~ 50% off input/output | — | — | 24-hour async SLA |
| Model | Input (≤200K) | Cached input | Output (≤200K) | Long-context (>200K) |
|---|---|---|---|---|
gemini-3.1-pro |
$2.00 | $0.20 | $12.00 | $4 / $18 |
gemini-3-pro |
$2.00 | $0.20 | $12.00 | $4 / $18 |
gemini-3-flash |
$0.50 | $0.05 | $3.00 | $1 / $6 |
gemini-2.5-pro |
$1.25 | $0.125 | $10.00 | $2.50 / $15 |
gemini-2.5-flash |
$0.075 | $0.02 | $0.30 | $0.15 / $0.60 |
| Imagen 3 (image gen) | — | — | $0.039/image | 1,290 output tokens per 1024×1024 |
| Veo 3 (video gen) | — | — | $0.40/sec | 1080p |
| Google AI Studio free tier | $0 | — | $0 | 100 req/day, all models |
Open-weight hosts (DeepSeek, Qwen, Llama)
Where self-hosting or low-cost hosted inference beats the frontier on price, the rate is noted. All prices are publicly listed hosted rates per 1M tokens as of July 2026; on-prem cost is roughly the GPU-hour equivalent.
| Model | Input | Output | Provider |
|---|---|---|---|
DeepSeek-V3 |
$0.27 | $1.10 | DeepSeek |
DeepSeek-R1 (reasoning) |
$0.55 | $2.19 | DeepSeek |
Qwen-3-Max |
$0.40 | $1.20 | Alibaba Cloud |
Qwen-3-235B |
$0.20 | $0.60 | Alibaba Cloud / Together |
Llama-4-405B |
$2.00 | $2.00 | Together / Fireworks |
Llama-4-70B |
$0.20 | $0.20 | Together / Fireworks |
What’s new in July 2026
- GPT-5.5 launch. OpenAI shipped
gpt-5.5at $5/$30 with the 90% caching discount in place.gpt-5.5-proat $30/$180 is the top of the OpenAI stack. Long-context doubles above 200K. - Claude 4.6 / 4.7 stable. Opus 4.7 is the flagship at $5/$25; Sonnet 4.6 ships on an introductory $2/$10 rate that expires on August 31, 2026. After that date, Sonnet 4.6 steps up to $3/$15.
- Gemini 3.1 launch + Flash-Lite deprecation. Google launched Gemini 3.1 Pro on June 15 and retired
gemini-flash-liteon June 1. Current cheapest paid Gemini model isgemini-2.5-flashat $0.075/$0.30. - Realtime price cuts.
gpt-realtime-2translation dropped from $0.05/min to $0.034/min in late June.whisperis now $0.017/min — usable for sub-$0.50 per-hour voice agents. - Long-context tiers tightened. All three frontier vendors now have explicit long-context (>200K) tiers that 2x the rate. Build the threshold into your pipeline before pushing long-context jobs to production.
- Open-weight competition intensifying.
Qwen-3-235Bat $0.20/$0.60 andLlama-4-70Bat $0.20/$0.20 are now genuinely production-viable for non-frontier workloads. Worth a routing decision per agent task.
Reading the rate cards correctly
A rate card is not a bill. Four things determine what you actually pay, and the August refresh of this hub audits all four on every model.
1. Cached vs uncached input. Raw rates are the worst case. With caching enabled (5-min default, extendable to 1 hour on most models), input costs drop 90%. If you’re running an agent with a system prompt above ~2K tokens and you don’t cache, you’re leaving roughly 70% of your input bill on the table. All three frontier vendors offer similar caching mechanics but bill writes differently — OpenAI and Anthropic charge ~1.25× input for the first cache write, Google charges near-parity.
2. Long-context tier (>200K). All three frontier vendors now have explicit long-context tiers that roughly double the rate past 200K. Audit your agent’s context sizes; a single over-threshold call can silently double a month’s bills. For multi-document analysis that genuinely needs 200K+, Gemini 3.1 Pro at $4/$18 is the cheapest option.
3. Output ratio. Output is 2–6x more expensive than input on every vendor. Workflows that produce long generations (research memos, full code, multi-step agent traces) should budget for output cost dominating the bill, not input. Reducing average output length by 30% typically drops the bill 20–25%.
4. Tool-use and search overhead. Tool-call schemas and grounding calls (web search, code execution) bill as tokens or per-1K-queries. Anthropic’s web search is $10 per 1,000 queries; Google’s grounding is $35 per 1,000 — the rates are inverted relative to what you’d expect. Track these line items separately; they often hide the actual margin.
How to use this hub — vendor decision tree
Picking the right API in 2026 is a routing problem, not a single-vendor decision. The following tree is what I’d recommend as a default routing layer for a solo operator or small agency in July 2026.
- Is the workload classification, tagging, or simple routing? Send to Google Gemini 2.5 Flash. Cheapest by 3x. Cache the prompt prefix; you should hit $0.02/M effective input.
- Does it involve ≥6 tool calls in a single turn? Send to Anthropic Sonnet 4.6 (or Opus 4.7 on the synthesis step). Tool-use adherence is the deciding factor.
- Does it need multimodal in one model — vision + tool use + image gen? Send to OpenAI
gpt-5.4orgpt-5.5. The cohesive tooling wins. - Is the dataset ≥500K tokens per call? Send to Anthropic Opus 4.7 or Google Gemini 3.1 Pro. Both maintain quality past 200K; OpenAI’s long-context tier is 2x and quality lags.
- Is it realtime voice? Send to OpenAI
gpt-realtime-2. Lowest per-minute cost; production-ready SLA. - Is it an open-weight model and you can host it? Self-host
Qwen-3-235Bon H100s for $0.20/M input equivalent. Worth it at 100M+ tokens/month. - Is it a prototype on a budget? Send to Google AI Studio free tier. Ship and graduate.
Subscribe to updates
This hub refreshes on the 1st of every month with the latest rates and any new model launches. Subscribe to the AI Pricing Watch newsletter for:
- Monthly rate-card diffs across OpenAI, Anthropic, Google, and the open-weight hosts.
- Routing-playbook updates when a major rate change ships.
- Quarterly reviews of the vendor decision tree.
- One worked bill per month on a real production workload.
[Newsletter signup form placeholder]
Related reads
- OpenAI API Pricing 2026: Every Model, Every Rate — the per-model deep dive for OpenAI with three real-workload bills.
- Anthropic Claude API Pricing 2026: Opus, Sonnet, Haiku — Claude rate card, Aug 31 deadline detail, and Anthropic’s prompt-caching math.
- Gemini API Pricing 2026: Is Google the Cheapest? — full Gemini rate card including the 200K tier, free tier, Imagen, and Veo.
- OpenAI vs Anthropic vs Google API Pricing 2026 — head-to-head comparison with four worked bills per vendor.
FAQ
How often does this hub update?
The 1st of every month. Per-vendor pages update on the same cadence. Last refresh was July 3, 2026.
How do you verify the rates?
Every rate is cross-checked against the public pricing dashboard of the corresponding vendor on the refresh date, plus a manual spot-check on the per-vendor deep-dive pages. Vendor dashboards can drift; if you find a discrepancy, drop us a note via the subscribe form.
Which vendor is cheapest overall?
Google, on raw token rates. gemini-2.5-flash at $0.075/M is the cheapest current model on the market. For cached rates, Gemini 2.5 Flash and OpenAI gpt-5.4-nano are tied at $0.02/M input.
What’s the most important rate change to watch?
Anthropic’s Aug 31, 2026 introductory rate expiry on Sonnet 4.6 (from $2/$10 to $3/$15). Plan budgets at the post-Aug 31 rate.
Should I start with one vendor or multi-vendor?
Build the routing abstraction up front. The multi-vendor agent saves 40–60% on most production workloads and you don’t want to refactor later.
Which vendor has the best rate for code review?
Google Gemini 3.1 Pro with caching at $4/$18 long-context input is the cheapest 200K-tier option; Anthropic Sonnet 4.6 with caching wins on tool-use adherence if you’re running multi-file refactor suggestions. OpenAI gpt-5.4 sits in between and is the strongest middle-ground pick if you don’t want to maintain two vendor integrations.
Which vendor has the best rate for image generation?
OpenAI gpt-image-1.5 medium at roughly $0.02–$0.06 per 1024×1024 image is the cheapest at standard sizes. Imagen 3 at $0.039/image is competitive at 1024×1024 but scales worse at 2048×2048.
What about embeddings costs?
OpenAI text-embedding-3-small at $0.02/M is the cheapest from a paid vendor. Google’s text-embedding-004 at $0.025/M is comparable. Anthropic doesn’t ship a first-party embedding API; use Voyage or OpenAI’s embedding API on Anthropic workloads.
Do these rates include prompt caching?
Yes — the cached-input rates are shown in the tables. Discounts are roughly 90% off raw input on every vendor that supports caching.
Where do I report a rate discrepancy?
Email the address on the subscribe form with the model name, vendor, and the rate you saw versus the rate we published. The page is corrected within 24 hours and an annotation is added to the next monthly refresh. Note: this page was last cross-checked against vendor dashboards on July 3, 2026.
How do I know if my billing matches your rate card?
Pull the last 30 days of usage from your vendor dashboard, total input and output tokens separately, multiply by the rate-table column matching your model and tier, and compare to your invoice. The five-line diff usually reveals either (a) a caching miss, (b) a long-context call you didn’t realise was over 200K, or (c) a per-call webhook or grounding charge that wasn’t on your radar.
Verdict
If you’re a solo operator or small agency owner, the July 2026 takeaway is: routing beats vendor loyalty. Multi-vendor setups that send classification to Gemini Flash, tool loops to Sonnet 4.6, and reasoning-heavy synthesis to Opus 4.7 or gpt-5.5 are 40–60% cheaper than single-vendor stacks — without measurable quality loss. Bookmark this page. Refresh it on the 1st of every month. Subscribe to the newsletter if you want the diff delivered. The next planned refresh lands on August 1, 2026, and will include the post-Aug 31 Sonnet 4.6 rate and any Q3 model launches from the frontier vendors.
If you’re a founder with a startup idea and no credit card, the takeaway is simpler: ship on Google AI Studio’s free tier. It still works in 2026, it covers all four flagship models, and it lets you validate a product with $0 of API spend. The moment volume justifies paying, you can graduate to paid Gemini without losing your work. OpenAI and Anthropic both killed their free tiers in 2025 and don’t have a credible re-launch plan before 2027. For cheap prototypes in July 2026, Google’s free tier is the only game in town.
This hub is updated monthly. Last refresh: July 3, 2026. Next refresh: August 1, 2026. Per-vendor deep dives updated in lockstep.
How we score: read the methodology
