LLM API Pricing 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell, AI and Machine Learning | Last updated: September 5, 2026 | License: CC BY 4.0
Prices verified on September 5, 2026. Every rate on this page was read directly from the vendor’s own published pricing documentation on that date. No aggregator, no third-party tracker, no cached figure.
OpenAI’s GPT-6 Astra costs $10.00 per million input tokens and $50.00 per million output tokens. Anthropic’s Claude Fable 5.1 costs $10.00 and $50.00. Those are not approximations of each other. They are the same two numbers, published by two competing labs, on two rate cards that otherwise share almost nothing structurally. The convergence at the top of the market is the single most under-reported fact in AI procurement right now, and it is the reason a rate card comparison that stops at the headline price tells you nothing useful.
Quick Answer: What Does an LLM API Cost in September 2026?
Across the twenty current production models Axis Intelligence Research price-checked against vendor documentation on September 5, 2026, published input rates run from $0.10 to $10.00 per million tokens and output rates from $0.40 to $50.00. According to Axis Intelligence Research, the median cost of running one thousand standardized agentic sessions across those twenty models is $112.50, with a floor of $11.00 (Gemini 2.5 Flash-Lite) and a ceiling of $1,250.00 (GPT-6 Astra and Claude Fable 5.1, tied) — a 113.6x spread for identical token volume. The headline rate is only one of four to six prices each vendor publishes for the same model.
Key Findings
- The two frontier flagships are priced identically. OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 both list at $10.00 input and $50.00 output per million tokens as of September 5, 2026, per each vendor’s own pricing documentation.
- According to Axis Intelligence Research, one model can carry an eight-fold spread in its own published price. GPT-6 Astra’s Rate Card Dispersion (RCD™) on input tokens is 8.0x — $5.00 per million on the Batch tier at short context, $40.00 on Fast mode at long context, both listed on the same OpenAI page.
- Crossing a context threshold by a single token can double the bill for the entire request. Google prices Gemini 3.1 Pro at $2.00 input for prompts at or below 200,000 tokens and $4.00 above it; xAI applies the same cliff to Grok 4.6. Axis Intelligence Research calculates that a 200,001-token request costs 1.97x a 200,000-token request on Gemini 3.1 Pro and exactly 2.00x on Grok 4.6.
- Anthropic is the only major vendor publishing no long-context surcharge. Its documentation states that Claude 4.6 and later models bill a 900,000-token request at the same per-token rate as a 9,000-token request.
- According to Axis Intelligence Research, routing plus caching beats vendor selection by a wide margin. A 73,000-session annual workload costs $91,250 on Claude Fable 5.1 alone and $20,727 on a routed mix of Sonnet 5, Opus 5 and Fable 5.1 with 90 percent prompt-cache hit rates — a 77.3 percent reduction achieved without leaving the vendor.
How Much Does Each LLM API Cost Per Million Tokens?
Here is the standard synchronous, uncached, short-context rate card for every current production text model Axis Intelligence Research verified. This is the number every pricing page leads with, and the number that is least useful on its own.
| Model | Provider | Input ($/MTok) | Output ($/MTok) | Output:input ratio | Verified | Source |
|---|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | 10.00 | 50.00 | 5.0x | 2026-09-05 | OpenAI API pricing |
| Claude Fable 5.1 | Anthropic | 10.00 | 50.00 | 5.0x | 2026-09-05 | Anthropic pricing docs |
| Claude Opus 5 | Anthropic | 5.00 | 25.00 | 5.0x | 2026-09-05 | Anthropic pricing docs |
| GPT-5.6 Sol | OpenAI | 4.00 | 20.00 | 5.0x | 2026-09-05 | OpenAI API pricing |
| Gemini 3.1 Pro | 2.00 | 12.00 | 6.0x | 2026-09-05 | Gemini API pricing | |
| GPT-5.6 Terra | OpenAI | 2.00 | 12.00 | 6.0x | 2026-09-05 | OpenAI API pricing |
| Claude Sonnet 5 | Anthropic | 2.00 | 10.00 | 5.0x | 2026-09-05 | Anthropic pricing docs |
| Grok 4.6 | xAI | 2.00 | 6.00 | 3.0x | 2026-09-05 | xAI API pricing |
| DeepSeek V4-Pro (peak) | DeepSeek | 1.32 | 3.96 | 3.0x | 2026-09-05 | DeepSeek models and pricing |
| Grok 4.3 | xAI | 1.25 | 2.50 | 2.0x | 2026-09-05 | xAI API pricing |
| Claude Haiku 4.5 | Anthropic | 1.00 | 5.00 | 5.0x | 2026-09-05 | Anthropic pricing docs |
| Grok Build 0.1 | xAI | 1.00 | 2.00 | 2.0x | 2026-09-05 | xAI API pricing |
| Gemini 3.8 Flash | 0.75 | 3.75 | 5.0x | 2026-09-05 | Gemini API pricing | |
| DeepSeek V4-Pro (off-peak) | DeepSeek | 0.66 | 1.98 | 3.0x | 2026-09-05 | DeepSeek models and pricing |
| DeepSeek V4-Flash (peak) | DeepSeek | 0.44 | 1.32 | 3.0x | 2026-09-05 | DeepSeek models and pricing |
| Gemini 3.5 Flash-Lite | 0.30 | 2.50 | 8.3x | 2026-09-05 | Gemini API pricing | |
| Gemini 3.1 Flash-Lite | 0.25 | 1.50 | 6.0x | 2026-09-05 | Gemini API pricing | |
| DeepSeek V4-Flash (off-peak) | DeepSeek | 0.22 | 0.66 | 3.0x | 2026-09-05 | DeepSeek models and pricing |
| GPT-5.6 Luna | OpenAI | 0.20 | 1.20 | 6.0x | 2026-09-05 | OpenAI API pricing |
| Gemini 2.5 Flash-Lite | 0.10 | 0.40 | 4.0x | 2026-09-05 | Gemini API pricing |
Source: OpenAI, Anthropic, Google, xAI and DeepSeek published pricing documentation, all retrieved September 5, 2026. Rates are USD per million tokens, standard synchronous tier, uncached, short context where the vendor tiers by context length.
The output-to-input column is the one to sit with. Most comparisons average input and output into a single blended figure and lose the structure entirely. xAI prices Grok 4.3 and Grok Build 0.1 at a 2:1 output multiple. Google prices Gemini 3.5 Flash-Lite at 8.3:1. That is a four-fold difference in how the same $1 of spend behaves depending on whether your workload writes or reads — and it is invisible on every blended chart in the category.
Sarah Mitchell’s read: the ratio is a product decision disguised as a price. A 2:1 multiple says the vendor wants you generating long outputs on this model. An 8:1 multiple says the vendor built a cheap classifier and would rather you did not use it as a writer. When you pick a model by input price and then discover the bill, the ratio is usually what happened.
The Rate Card Dispersion Metric: One Model, Eight Prices
Every serious pricing comparison in this category has the same blind spot. It treats “the price of GPT-6 Astra” as a scalar. It is not. It is a range, and the range is published on the vendor’s own page.
Axis Intelligence Research created RCD™ — Rate Card Dispersion to measure it.
Formula: RCD = highest published self-serve list price for a model ÷ lowest published self-serve list price for the same model, computed separately for input tokens and output tokens.
Inputs: every service tier and context tier that the vendor publishes on its own developer pricing documentation for that model — standard, batch, flex, priority, fast mode, short context, long context, peak and off-peak.
Scope: self-serve list prices only. Cache reads and cache writes are excluded because they price a different unit of work. Negotiated enterprise rates, volume discounts, marketplace resale and data-residency multipliers are excluded because they are not published.
| Model | Provider | Input range ($/MTok) | Input RCD™ | Output range ($/MTok) | Output RCD™ |
|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | 5.00 to 40.00 | 8.0x | 25.00 to 150.00 | 6.0x |
| GPT-5.6 Sol | OpenAI | 2.00 to 16.00 | 8.0x | 10.00 to 60.00 | 6.0x |
| Gemini 3.1 Pro | 1.00 to 7.20 | 7.2x | 6.00 to 32.40 | 5.4x | |
| Grok 4.6 | xAI | 2.00 to 8.00 | 4.0x | 6.00 to 24.00 | 4.0x |
| Claude Opus 5 | Anthropic | 2.50 to 10.00 | 4.0x | 12.50 to 50.00 | 4.0x |
| Gemini 3.8 Flash | 0.375 to 1.35 | 3.6x | 1.875 to 6.75 | 3.6x | |
| Claude Fable 5.1 | Anthropic | 5.00 to 10.00 | 2.0x | 25.00 to 50.00 | 2.0x |
| DeepSeek V4-Flash | DeepSeek | 0.22 to 0.44 | 2.0x | 0.66 to 1.32 | 2.0x |
Source: Axis Intelligence Research calculation from vendor pricing documentation retrieved September 5, 2026. Arithmetic verified September 5, 2026.
According to Axis Intelligence Research, GPT-6 Astra carries the widest published dispersion of any current frontier model at 8.0x on input, driven by the interaction of two independent multipliers OpenAI applies to the same model: a 50 percent Batch discount and a Fast mode premium, layered over a long-context tier that doubles input and adds 50 percent to output. Claude Fable 5.1 and DeepSeek V4-Flash sit at the opposite end at 2.0x, for opposite reasons. Anthropic publishes no long-context tier and no Fast mode for Fable, leaving Batch as the only modifier. DeepSeek publishes no batch tier at all, leaving its peak and off-peak clock windows as the only variation.
A high RCD is not a criticism. It means the vendor has given you levers. A low RCD means the price you see is close to the price you pay. What it does mean is that any comparison quoting one number per model is comparing points chosen arbitrarily from ranges of very different widths.
RCD™ is published under Creative Commons Attribution 4.0 International (CC BY 4.0). To cite: Axis Intelligence Research Rate Card Dispersion, September 2026 reading, axis-intelligence.com/llm-api-pricing/.
What Does One Agentic Session Actually Cost?
Per-million-token pricing is a wholesale unit. Nobody buys a million tokens. Teams buy sessions, requests and completed jobs, and the only way to compare vendors honestly is to run identical work through every rate card.
Axis Intelligence Research uses CPAS™ — Cost Per Agentic Session for this.
Formula: CPAS = (50,000 ÷ 1,000,000 × input price) + (15,000 ÷ 1,000,000 × output price), reported per 1,000 sessions.
Inputs: the 50,000 input / 15,000 output token profile is not an Axis invention. It is the one-hour coding session Anthropic publishes as a worked billing example in its Claude Managed Agents pricing documentation. Axis Intelligence Research adopts it unchanged as a cross-vendor constant so that every model is priced on identical token volume rather than on each vendor’s flattering example.
Scope: standard synchronous tier, uncached, short context. Session runtime fees, tool-invocation fees, storage and reasoning-token expansion are excluded from the base reading and handled separately below.
| Model | Provider | CPAS™ per 1,000 sessions | Annual at 73,000 sessions |
|---|---|---|---|
| GPT-6 Astra | OpenAI | $1,250.00 | $91,250 |
| Claude Fable 5.1 | Anthropic | $1,250.00 | $91,250 |
| Claude Opus 5 | Anthropic | $625.00 | $45,625 |
| GPT-5.6 Sol | OpenAI | $500.00 | $36,500 |
| Gemini 3.1 Pro | $280.00 | $20,440 | |
| GPT-5.6 Terra | OpenAI | $280.00 | $20,440 |
| Claude Sonnet 5 | Anthropic | $250.00 | $18,250 |
| Grok 4.6 | xAI | $190.00 | $13,870 |
| DeepSeek V4-Pro (peak) | DeepSeek | $125.40 | $9,154 |
| Claude Haiku 4.5 | Anthropic | $125.00 | $9,125 |
| Grok 4.3 | xAI | $100.00 | $7,300 |
| Gemini 3.8 Flash | $93.75 | $6,844 | |
| Grok Build 0.1 | xAI | $80.00 | $5,840 |
| DeepSeek V4-Pro (off-peak) | DeepSeek | $62.70 | $4,577 |
| Gemini 3.5 Flash-Lite | $52.50 | $3,833 | |
| DeepSeek V4-Flash (peak) | DeepSeek | $41.80 | $3,051 |
| Gemini 3.1 Flash-Lite | $35.00 | $2,555 | |
| GPT-5.6 Luna | OpenAI | $28.00 | $2,044 |
| DeepSeek V4-Flash (off-peak) | DeepSeek | $20.90 | $1,526 |
| Gemini 2.5 Flash-Lite | $11.00 | $803 |
Source: Axis Intelligence Research calculation from vendor pricing documentation retrieved September 5, 2026. Annual column assumes 200 sessions per day. Arithmetic verified September 5, 2026.
Three things fall out of that table that the per-token view hides.
Grok 4.6 and Claude Sonnet 5 list at the identical $2.00 input price, and Grok 4.6 lands 24 percent cheaper per session because its output multiple is 3:1 rather than 5:1. Gemini 3.5 Flash-Lite costs more per session than Gemini 3.1 Flash-Lite despite both being Lite-tier models, because Google raised the output rate from $1.50 to $2.50 across that generation while raising input only from $0.25 to $0.30. And DeepSeek V4-Pro at peak hours costs almost exactly what Claude Haiku 4.5 costs around the clock, which is a very different sentence from the one that gets written about DeepSeek pricing.
CPAS™ is published under CC BY 4.0. To cite: Axis Intelligence Research Cost Per Agentic Session, September 2026 reading, axis-intelligence.com/llm-api-pricing/.
The Fully Loaded Number: Long Context, Fast Mode and the Threshold Cliff
The base CPAS reading assumes a 50,000-token prompt. Real agentic workloads do not stay there. Once a request crosses a vendor’s context threshold, the pricing model changes shape, and it changes retroactively across the whole request.
xAI states the rule in its own documentation without hedging: models with long-context pricing bill the long-context rates for all tokens in a request once the prompt reaches the threshold. Google applies the same structure to Gemini 3.1 Pro at 200,000 tokens. According to Axis Intelligence Research, that makes a 200,001-token request cost 2.00x a 200,000-token request on Grok 4.6 and 1.97x on Gemini 3.1 Pro, with the difference arising because Google raises output by 50 percent while xAI doubles it. Sending 0.0005 percent more context can cost 100 percent more money.
Here is the same 1,000-session workload with a 250,000-token prompt instead of a 50,000-token one.
| Model | Provider | Applicable tier | Long-context CPAS™ per 1,000 | Short-context CPAS™ | Multiple |
|---|---|---|---|---|---|
| GPT-6 Astra | OpenAI | Long context | $6,125 | $1,250 | 4.9x |
| Claude Fable 5.1 | Anthropic | Flat | $3,250 | $1,250 | 2.6x |
| Claude Opus 5 | Anthropic | Flat | $1,625 | $625 | 2.6x |
| Gemini 3.1 Pro | Above 200k | $1,270 | $280 | 4.5x | |
| Grok 4.6 | xAI | Above 200k | $1,180 | $190 | 6.2x |
| DeepSeek V4-Flash | DeepSeek | Flat off-peak | $65 | $20.90 | 3.1x |
Source: Axis Intelligence Research calculation from vendor pricing documentation retrieved September 5, 2026. Long-context profile is 250,000 input and 15,000 output tokens per session.
The ordering changes. At 50,000 tokens, Grok 4.6 is 32 percent cheaper per session than Gemini 3.1 Pro. At 250,000 tokens it is only 7 percent cheaper, because the same threshold that doubles Grok’s input rate raises Gemini’s output by half rather than double. Meanwhile Claude Opus 5 goes from 3.3x the cost of Grok 4.6 at short context to 1.4x at long context, purely because Anthropic does not tier by prompt length.
Then there is speed. Anthropic charges $10.00 input and $50.00 output for Fast mode on Opus 5 and Opus 4.8 — exactly double standard, and exactly the Fable 5.1 rate. OpenAI’s Fast mode runs 2x standard. xAI’s Priority Processing runs 2x standard and only bills at the priority rate when the response confirms the priority service tier. Google’s Priority tier on Gemini 3.1 Pro runs 1.8x standard. Four vendors, four different speed premiums, one shared idea that latency is a billable product.
Sarah Mitchell’s read: the threshold cliff is the most expensive line of documentation any of these vendors has written, and it is a footnote on all of them. If your retrieval layer sizes context dynamically and your prompt sometimes lands at 205,000 tokens, you are paying the long-context rate on a request that a slightly better chunking policy would have kept under the line. That is a compaction problem being paid for out of the inference budget.
What Moves the Price
Six mechanisms account for essentially every gap between a rate card and an invoice.
Prompt caching, and how unevenly it is priced. Anthropic charges 10 percent of base input on a cache hit for most models and 2.5 percent on Fable 5.1 and Mythos 5.1, which is the lowest published cache-read multiplier of any vendor. xAI publishes $0.50 per million cached input on Grok 4.6, a 75 percent discount off standard rather than the 90 percent the others offer, and $0.20 on Grok 4.3. Google prices context caching per model plus a storage fee per million tokens per hour — $4.50 an hour on Gemini 3.1 Pro, which turns caching into a decision with a clock on it. DeepSeek’s cache-hit rate on V4-Flash is $0.007 off-peak, roughly 3 percent of its cache-miss rate.
According to Axis Intelligence Research, at a 90 percent cache-hit rate the effective input price on Claude Fable 5.1 falls from $10.00 to $1.225 per million tokens and CPAS drops from $1,250.00 to $811.25 per thousand sessions. On Claude Opus 5 the same hit rate takes effective input from $5.00 to $0.95 and CPAS from $625.00 to $422.50.
Batch, and the vendors who do not offer it. OpenAI, Anthropic and Google all discount 50 percent for asynchronous work. xAI does not offer a batch discount on Grok 4.6 at all; its published batch discount is 20 percent and applies only to Grok 4.3 and the Grok 4.20 family. Any bulk-processing comparison that assumes a universal 50 percent batch floor is wrong on xAI by a factor of two.
Time of day. DeepSeek moved to peak and off-peak billing effective 16:00 UTC on August 16, 2026, with off-peak set at half the peak rate and peak hours running 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Every other hour, including all weekend hours, is off-peak. That is 48 of every 168 hours guaranteed cheap. No other major vendor prices by clock.
Data residency. Anthropic applies a 1.1x multiplier across all token categories when inference is pinned to US-only geography on Claude 4.6 and later. OpenAI charges a 10 percent uplift on regional processing endpoints for models released on or after March 5, 2026. Partner-operated clouds price regionally on their own terms. Sovereignty is a line item now, and it is roughly a dime on the dollar.
Tools, which are not tokens. Anthropic charges $10 per 1,000 web searches and $0.08 per session-hour of Managed Agents runtime. OpenAI charges $10 per 1,000 web search calls and $2.50 per 1,000 file-search calls. xAI charges $5 per 1,000 calls for web search, X search and code execution, $10 per 1,000 for file attachments, and $2.50 per 1,000 for collections search, plus storage at $0.025 per GiB per day and $0.20 per GiB downloaded. Google charges $14 per 1,000 grounded search requests on the Gemini 3.x line after 5,000 free per month. An agent that calls three tools per step accrues cost on a second meter entirely.
Scheduled price changes. Google’s Gemini 3.6, 3.7 and 3.8 Flash models carry introductory pricing of $0.75 and $3.75 through December 31, 2026, moving to $1.50 and $7.50 on January 1, 2027. According to Axis Intelligence Research, that step-up exactly doubles Gemini 3.8 Flash’s CPAS from $93.75 to $187.50 per thousand sessions, taking a 73,000-session annual workload from $6,844 to $13,688 with no change in usage. Anthropic moved the other way on Sonnet 5, confirming that the $2 and $10 introductory rate is now the standard price and that the scheduled increase to $3 and $15 on September 1, 2026 will not occur.
Do Published Rates Reflect What Large Buyers Pay?
Not always, and the clearest public evidence comes from federal procurement rather than from any vendor.
Under the General Services Administration’s OneGov strategy, Anthropic agreed to provide Claude for Enterprise and Claude for Government to eligible federal agencies for $1 per agency for up to a full year, and Google agreed to provide its Gemini for Government suite at $0.47 per agency. OpenAI struck a comparable $1-per-agency arrangement for ChatGPT Enterprise.
Two precisions matter here, and most coverage skips both. These are seat-based enterprise product agreements, not per-token API rates — an agency on a OneGov deal is not getting GPT-6 Astra inference for $1 per million tokens. And they are time-boxed introductory terms rather than standing prices. What they do establish, from a primary government source rather than a vendor blog, is that the published rate card is a starting position for buyers of sufficient scale. Every vendor in this comparison directs high-volume customers to sales for custom pricing; Anthropic, OpenAI, Google and xAI all say so on the same pages that carry the list rates.
The practical reading for a commercial buyer: if your annual committed spend is in six figures, the rate card is the number you should be negotiating away from, not budgeting to.
A Calculation Axis Intelligence Research Declined to Publish
This comparison originally included Mistral AI as the European entry. It does not, and the reason is worth stating.
Mistral’s first-party rate card could not be retrieved directly during the September 5, 2026 production session. Secondary sources published within the last thirty days disagree materially on the flagship: some list Mistral Large 3 at $0.50 input and $1.50 output per million tokens, others at $2.00 and $6.00 — a four-fold divergence on the same model in the same month. Publishing a CPAS reading built on either figure would have produced a precise-looking number resting on an unresolved conflict, and it would have placed Mistral either second-cheapest or mid-table in the ranking depending entirely on which unverified source we picked.
The row is retracted rather than estimated. It is recorded in the accompanying dataset with a data type of retracted so that anyone reproducing this analysis can see exactly where we stopped. Mistral will be added at the next refresh if the first-party rate card is reachable.
Which LLM API Should You Actually Buy?
Verdicts here follow the arithmetic. Axis Intelligence Research holds no commercial relationship with any vendor named on this page, and nothing on it is affiliate-placed.
For high-volume classification, extraction and routing: Gemini 2.5 Flash-Lite at $11.00 CPAS per thousand sessions is the cheapest published option in this set by a factor of two over its nearest rival, and its 4:1 output multiple is more forgiving than the 6:1 and 8.3:1 ratios elsewhere in Google’s Lite range. DeepSeek V4-Flash off-peak at $20.90 is competitive if your workload can be scheduled and your data can sit outside the EU and US.
For general production work: Claude Sonnet 5 at $250.00 and Grok 4.6 at $190.00 anchor this band. Grok is cheaper per session on the standard tier; Sonnet is cheaper once you cache, because Anthropic’s 10 percent cache-read multiplier beats xAI’s 25 percent, and Sonnet has a 50 percent batch tier where Grok 4.6 has none. If your workload is synchronous and uncached, take Grok. If it caches well or batches, take Sonnet.
For long-context agentic work: Anthropic’s flat pricing across the full context window is a structural advantage that grows with prompt size. Claude Opus 5 at $1,625 per thousand long-context sessions costs less than Gemini 3.1 Pro did before the threshold and only 1.4x what Grok 4.6 costs after it.
For frontier-only tasks: GPT-6 Astra and Claude Fable 5.1 cost the same. Choose on evaluations, not price, because there is no price to choose on. Fable 5.1’s 2.5 percent cache-read rate is the one published economic differentiator between them, and it is worth roughly 35 percent of the bill on a cache-heavy agentic workload.
The lever that beats all of the above: According to Axis Intelligence Research, routing a 73,000-session annual workload 70 percent to Sonnet 5, 20 percent to Opus 5 and 10 percent to Fable 5.1 costs $31,025 against $91,250 for running everything on Fable 5.1 — a 66.0 percent reduction. Layering 90 percent prompt-cache hit rates on top brings it to $20,727, or 77.3 percent below the single-model baseline, entirely within one vendor. No cross-vendor migration in this comparison delivers a saving of that size.
For the broader market context behind these rates — how frontier pricing has moved since GPT-4 and why total enterprise spend rises while unit prices fall — see the Axis Intelligence Research analysis of AI inference cost, which owns the LLMflation Index™ time series. For launch dates and model lineage behind the names in the table above, the AI Model Release Tracker carries the log. Company-level financial context sits in the OpenAI and Anthropic statistics pages, and the infrastructure spending underneath all of it is tracked in the AI Capex Tracker.
Methodology
Collection. Every price in this article was read from the vendor’s own published pricing documentation during the production session of September 5, 2026: OpenAI at developers.openai.com/api/docs/pricing, Anthropic at platform.claude.com/docs/en/about-claude/pricing, Google at ai.google.dev/gemini-api/docs/pricing, xAI at docs.x.ai/developers/pricing, and DeepSeek at api-docs.deepseek.com/quick_start/pricing. No third-party pricing tracker, aggregator or comparison site supplied any figure in the tables above. Where a vendor publishes both an introductory and a scheduled standard rate, both are recorded and the applicable date stated.
CPAS™ calculation. CPAS = (50,000 ÷ 1,000,000 × input price) + (15,000 ÷ 1,000,000 × output price). The session profile is Anthropic’s published one-hour coding session worked example, adopted unchanged as a cross-vendor constant. Readings use the standard synchronous tier, uncached, at short-context rates where a vendor tiers by prompt length. The long-context variant substitutes a 250,000-token prompt and applies each vendor’s above-threshold rate. Annual figures assume 200 sessions per day across 365 days.
RCD™ calculation. RCD = highest published self-serve list price ÷ lowest, per token type, across every service tier and context tier the vendor publishes for that model. Cache reads and writes are excluded as a different unit of work. Negotiated, marketplace and residency-adjusted rates are excluded as unpublished.
Exclusions. Reasoning-token expansion is not modeled; Google bills thinking tokens at output rates and other vendors do the same, so real output volumes on reasoning models exceed the visible response. Tool-invocation fees, session-runtime fees, storage and egress are described in the price-drivers section but excluded from CPAS so that the metric compares token economics only. Retries, taxes, and fallback-provider routing are excluded.
Verification. All arithmetic was computed in Python and re-run against the source rate cards on September 5, 2026. Every numeric token in this article was matched in both directions against the accompanying dataset. One vendor row was retracted rather than estimated, and the reason is published above.
About This Dataset
Coverage: published API rate cards for 20 current production text models across five vendors, plus service-tier and context-tier variants, September 5, 2026 snapshot. Includes CPAS™ and RCD™ readings for every model.
Primary sources: OpenAI, Anthropic, Google, xAI and DeepSeek developer pricing documentation; US General Services Administration OneGov news releases.
Download: CC BY 4.0. Cite as: Axis Intelligence Research, LLM API Pricing 2026, axis-intelligence.com/llm-api-pricing/.
Questions Teams Ask Before They Sign
Why do two competing labs publish the exact same $10 and $50 rate?
Because frontier pricing has become a positioning signal rather than a cost pass-through. OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 both list at $10.00 input and $50.00 output per million tokens as of September 5, 2026. Neither vendor discloses inference cost, so the convergence cannot be read as a shared cost floor. What it does establish is a ceiling: $50 per million output tokens is currently the most any major lab will ask for a self-serve frontier model, and both have priced to it rather than through it. The practical consequence for buyers is that at the top tier the differentiation is entirely in modifiers — Fable 5.1’s 2.5 percent cache-read multiplier against Astra’s 50 percent batch discount and Fast mode option.
My prompt sits near 200,000 tokens. What does that actually cost me?
More than you think, because the threshold is a cliff rather than a slope on both Google and xAI. Both bill the higher rate across every token in the request once the prompt reaches the threshold, not just the tokens above it. Axis Intelligence Research calculates that a 200,001-token request on Grok 4.6 costs exactly 2.00x a 200,000-token request, and 1.97x on Gemini 3.1 Pro. If your context assembly is dynamic, the fix is a hard cap slightly below the threshold with compaction above it, not a bigger budget. Anthropic is the exception in this comparison: its documentation states the full 1M-token window bills at standard rates on Claude 4.6 and later.
Does the 50 percent batch discount apply everywhere?
No, and assuming it does is the most common budgeting error in this category. OpenAI, Anthropic and Google all publish a 50 percent batch discount on both input and output. xAI publishes a 20 percent batch discount that applies only to Grok 4.3 and the Grok 4.20 family — Grok 4.6, its current flagship, has no batch discount at all. DeepSeek publishes no batch tier, offering time-of-day pricing instead. If you are modeling a bulk pipeline on xAI’s flagship, use the standard rate.
How much does caching actually change the ranking?
Enough to reorder it. At a 90 percent cache-hit rate, Axis Intelligence Research calculates effective input prices of $1.225 per million on Claude Fable 5.1, $0.95 on Opus 5, $0.38 on Sonnet 5 and $0.19 on Haiku 4.5. That takes Sonnet 5’s CPAS from $250.00 to $169.00 per thousand sessions and moves it below Grok 4.6’s uncached $190.00, reversing the standard-tier ordering. Cache-read multipliers vary more than base rates do: 2.5 percent on Fable 5.1, 10 percent across the rest of the Claude line, 25 percent on Grok 4.6, and roughly 3 percent on DeepSeek V4-Flash. Google adds an hourly storage charge on top, which makes its cache a decision about dwell time rather than hit rate alone.
Is the cheapest model per token the cheapest model per job?
Only when the job is trivial. Per-token price ignores three multipliers that decide real bills: the output ratio, which varies from 2:1 to 8.3:1 across the models in this comparison; reasoning-token expansion, where internal thinking bills at output rates and is invisible in the response; and retry rate, since a model that fails twice at one-fifth the price costs more than one that succeeds once. CPAS™ addresses the first by holding token volume constant across vendors. The other two require your own evaluations — no rate card can tell you how often a given model gets your task right on the first attempt.
What should trigger a re-check of these numbers?
Any vendor pricing-page change, any new frontier model launch, and three dated events already on the calendar: Google’s Gemini 3.6, 3.7 and 3.8 Flash introductory rates expire December 31, 2026 and double on January 1, 2027; OpenAI’s promotional pricing on GPT-5.6 Sol is stated as available at least through November 21, 2026; and DeepSeek has signaled further pricing changes without publishing replacement rates. Axis Intelligence Research re-verifies every rate on this page before each refresh rather than carrying forward prior readings.
Citation Block
APA: Axis Intelligence Research & Mitchell, S. (2026, September 5). LLM API Pricing 2026: Every Frontier Rate Card, Normalized to One Number. Axis Intelligence Research. https://axis-intelligence.com/llm-api-pricing/
MLA: Axis Intelligence Research and Sarah Mitchell. “LLM API Pricing 2026: Every Frontier Rate Card, Normalized to One Number.” Axis Intelligence Research, 5 Sept. 2026, axis-intelligence.com/llm-api-pricing/.
Chicago: Axis Intelligence Research and Sarah Mitchell. “LLM API Pricing 2026: Every Frontier Rate Card, Normalized to One Number.” Axis Intelligence Research, September 5, 2026. https://axis-intelligence.com/llm-api-pricing/