AI API Pricing Compared: Every Frontier Model (Sept 2026)

Every price is US dollars per million tokens, standard tier, for prompts under 200K tokens unless noted. "Cached input" is the cache-read rate. "n/a" means the cached rate is not in the sources we verified, not that caching is unavailable.

Checked 23 September 2026. Sources: Claude rates from Anthropic's pricing documentation (platform.claude.com); OpenAI rates from OpenAI's announcements as reported by VentureBeat, Vellum and DataCamp; Gemini and Grok rates from Google's and xAI's published list prices. Full links in Sources.

Long context matters for pricing. All four current Claude models listed with 1M context (Opus 5.5, Opus 5, Fable 5.1 and Sonnet 5) charge the standard rate across the full 1M window, with no long-context premium. Gemini 3.1 Pro doubles input to $4 and raises output to $18 above 200K tokens. Grok 4.7 moves to $4 input, $1 cached and $12 output at 200K and above. If your prompts routinely pass 200K tokens, recalculate Gemini and Grok at those rates before comparing.

The rate card falls into four clear bands.

• Budget (under $1 input): GPT-6 Luna ($0.10/$0.50), Gemini 3.8 Flash ($0.75/$3.75 intro, $1.50/$7.50 from 2027) and Claude Haiku 4.5 ($1/$5). Luna is 7.5x cheaper than Flash on output and 10x cheaper than Haiku.
• Mid-tier ($2 input): GPT-6 Sol ($2/$10), Claude Sonnet 5 ($2/$10), Gemini 3.1 Pro ($2/$12) and Grok 4.7 ($2/$6). Input is identical across all four. Output is where they split, from Grok's $6 to Gemini Pro's $12.
• Flagship: Claude Opus 5.5 ($4/$20) and GPT-6 Astra ($10/$50). Astra costs 2.5x Opus 5.5 on both input and output, and 5x on cached input ($1.00 vs $0.20). Opus 5 ($5/$25) still lists but is superseded.
• Specialist: Claude Fable 5.1 ($10/$50). Same headline rate as Astra, but with the cheapest relative cache read on the card: $0.25, or 2.5% of its input price.

The mid-tier is now the most crowded band. Four vendors sit at $2 input. For the tier-by-tier quality picture, see our GPT-6 Sol vs Astra vs Luna breakdown and the September 2026 frontier benchmarks round-up.

Per-million rates are hard to reason about. So we priced three typical monthly workloads on all 11 current and recent models. These are AI Tools Review's own calculations, computed 23 September 2026 from the list prices above.

• A. Chatbot: 10M input tokens a month (80% cache hits) and 2M output tokens.
• B. Coding agent: 50M input tokens a month (90% cache hits) and 5M output tokens.
• C. Bulk extraction: 100M input tokens a month (no caching) and 10M output tokens.

Method: monthly token volume multiplied by the standard list price. The cached share is billed at the cache-read rate. Cache-write charges, batch discounts, fast mode and tool fees are excluded.

Checked 23 September 2026. Source: AI Tools Review workload cost model, from the list prices in the rate card above. *Gemini's cache-read price is not in our verified sources, so Gemini rows assume no cache discount. That overstates Gemini's cost on workloads A and B. Gemini 3.8 Flash is shown at its introductory rate; every Flash figure doubles at the post-intro rate from 1 January 2027.

Read this with one caveat in mind. These are per-token list costs. Models use different numbers of tokens for the same task, because tokenisers, verbosity and reasoning length all differ. Real cost per task can land well away from these figures, in either direction. Treat the table as a ranking of list prices applied to identical token counts, not a forecast of your invoice.

What the model shows:

• GPT-6 Sol and Claude Sonnet 5 are identical on all three workloads, because their list prices are identical.
• Opus 5.5 costs about 39-40% of GPT-6 Astra on every workload: $129 against $345 for the coding agent.
• Opus 5.5 is 22-25% cheaper than Opus 5 at list price. Anthropic's own "about 40% cheaper" figure also counts fewer tokens per task, which this model does not capture.
• Caching reorders the budget tier. Haiku 4.5 beats Gemini 3.8 Flash on the cached workloads A and B, but Flash wins the uncached bulk job ($112.50 vs $150).
• Gemini 3.1 Pro looks expensive on the coding agent ($160, more than Opus 5.5) only because we could not apply a cache discount. On uncached bulk extraction it is $320, close to Sol and Sonnet 5.

• GPT-6 Sol replaced GPT-5.6 Sol at half its promotional price: $4/$20 (standard $5/$30) became $2/$10, and cached input fell from $0.50 to $0.20.
• GPT-6 Luna replaced GPT-5.6 Luna: input halved from $0.20 to $0.10, output fell from $1.20 to $0.50, and cached input halved to $0.01. OpenAI says both new prices are permanent.
• Claude Opus 5.5 launched at 20% below Opus 5 on input and output ($4/$20 vs $5/$25). Cache reads fell 60%, from $0.50 to $0.20.

One earlier change also matters. Claude Sonnet 5's $2/$10 introductory price was made permanent on 11 August 2026. Anthropic had planned to raise it to $3/$15 on 1 September and cancelled the rise. That is why Sonnet 5 and GPT-6 Sol now sit at exactly the same price. The full Claude history is in our Claude API pricing guide, and the Sol and Luna launch is covered in GPT-6 Sol and Luna explained.

Prompt caching is the biggest lever. Cache reads are priced as a fraction of the input rate, and that fraction varies:

• Claude Opus 5.5: 0.05x input ($0.20 against $4), a 95% discount.
• Claude Fable 5.1: 0.025x input ($0.25 against $10), a 97.5% discount.
• Claude Opus 5, Sonnet 5 and Haiku 4.5: 0.1x input, a 90% discount.
• GPT-6 Astra, Sol and Luna: 90% off input.
• Grok 4.7: $0.50 against $2, a 75% discount.
• Gemini: not in our verified sources.

Cache writes cost extra. On Claude, a 5-minute write is $5 on Opus 5.5, $6.25 on Opus 5, $12.50 on Fable 5.1 and $2.50 on Sonnet 5. A 1-hour write is $8 on Opus 5.5 and $4 on Sonnet 5. GPT-6 Astra lists a $12.50 cache write. Our workload model leaves these out, so heavy cache churn will cost more than shown.

Batch halves Claude prices. Anthropic's batch rates are 50% of standard: Opus 5.5 $2/$10, Opus 5 $2.50/$12.50, Fable 5.1 $5/$25 and Sonnet 5 $1/$5. For our bulk-extraction workload, which does not need a real-time answer, that is the easiest saving available. Batch rates for OpenAI, Google and xAI are not in our verified sources, so we have not listed them.

Fast mode costs a premium. Claude Opus 5.5 fast mode is $8/$40, double the standard rate. Opus 5 fast mode is $10/$50. GPT-6 Astra's fast mode is 2x the price for up to 2.5x the speed. Use it only where latency earns money.

Watch the Gemini deadline. Gemini 3.8 Flash's $0.75/$3.75 rate runs through 31 December 2026. From then, $1.50/$7.50 puts its output price above Claude Haiku 4.5's $5. Budget for the higher rate on anything that runs into 2027.

A per-token rate only tells you half the bill. The other half is how many tokens a model spends to finish the job.

Grok 4.7 has the lowest mid-tier output price at $6. But independent testers describe it as slow and verbose, so its cost per task is higher than the per-token rate suggests. Its $62.50 coding-agent figure in our model assumes the same token count as every other model. In practice it is likely to use more.

Claude Opus 5.5 runs the other way. Anthropic cut its list price 20% but says typical workloads cost about 40% less than Opus 5, because Opus 5.5 uses fewer tokens per task. Anthropic also claims that at default effort it beats GPT-6 Astra on FrontierCode at about 20% of the cost per task, and matches Astra on Terminal-Bench 4.0 at about 40% of the cost. These are Anthropic's own figures, not independent measurements.

OpenAI makes the same argument for Sol. On AutomationBench, OpenAI reports GPT-6 Sol at 33.2% for $0.27 per task, and says that is 8.9x cheaper per task than Fable 5.1 and 11.1x cheaper than Opus 5. OpenAI did not compare against Opus 5.5, which launched about 90 minutes earlier. On the same benchmark, Anthropic reports Opus 5.5 at 40.0% but did not publish a comparable cost per task. For the head-to-head, see GPT-6 Sol vs Claude Opus 5.5.

The practical rule: shortlist on list price, then run 50-100 of your real prompts through two or three candidates and compare the invoices.

• Absolute lowest cost: GPT-6 Luna. $1.28 a month for our chatbot, $3.45 for the coding agent, $15 for bulk extraction.
• Cheapest Claude: Haiku 4.5 at $1/$5. It is still roughly 10x Luna on every workload we modelled.
• Bulk, uncached extraction: Luna, then Gemini 3.8 Flash at $112.50 while the intro rate lasts, then Haiku 4.5 at $150. On Claude, batch pricing halves the bill.
• Cheapest mid-tier by list price: Grok 4.7 ($62.50 coding agent), with the verbosity caveat above. GPT-6 Sol and Claude Sonnet 5 tie at $69.
• Cheapest flagship: Claude Opus 5.5. $129 on the coding agent against $345 for GPT-6 Astra.
• Long prompts over 200K tokens: Claude, which charges its standard rate across the full 1M window. Gemini 3.1 Pro and Grok 4.7 both step up above 200K.
• Heavily cached agent loops: Opus 5.5 among flagships, with cache reads at 5% of input.

• Do not choose GPT-6 Astra on cost. It is the most expensive model on the card, tied with Fable 5.1 on bulk extraction and the dearest on the other two workloads. Pick it only where it wins on quality for your task.
• Do not stay on Claude Opus 5. Opus 5.5 is cheaper on every rate and scores higher on every benchmark in Anthropic's launch table.
• Do not pick Claude Fable 5.1 for general work on price. It costs 2.5x Opus 5.5 on list price, and Opus 5.5 beats it on every benchmark Anthropic published. Its case is specialist work, not cost.
• Do not pick Grok 4.7 on the per-token rate alone. Tester reports of verbose output mean the per-task bill may erase its $6 output advantage.
• Do not lock a 2027 budget to Gemini 3.8 Flash's intro price. It doubles after 31 December 2026.
• Do not assume the cheapest tier is good enough. On OpenAI's own DeepSWE v1.1 numbers, Luna (66.6%) sits close to Sol (68.8%), but that is one benchmark from one vendor. Test your own tasks.

Subscription buyers face a different calculation. Claude Pro is $20 a month and Max is $100 or $200, unchanged by the Opus 5.5 launch. See Claude pricing for plan limits, and Claude vs ChatGPT vs Gemini vs Grok for the consumer comparison.

• Anthropic: Claude API pricing documentation
• Anthropic: Claude Opus 5.5 announcement
• OpenAI: Introducing GPT-6 Sol and Luna
• OpenAI: GPT-6 Astra
• VentureBeat: OpenAI releases GPT-6 Sol and Luna, slashing API costs 50% or more
• TechCrunch: OpenAI launches GPT-6 Sol and Luna
• Vellum: GPT-6 Sol and Luna benchmarks explained
• DataCamp: GPT-6 Astra
• Artificial Analysis: GPT-6 Sol

Gemini 3.8 Flash, Gemini 3.1 Pro and Grok 4.7 rates are Google's and xAI's published list prices as checked by AI Tools Review on 23 September 2026. Workload costs are AI Tools Review's own calculations.