The Frontier-Model API Price Floor: What Agents Cost Now

As of September 2026, the frontier "workhorse" model has a de facto list price: about $2 per million input tokens and $10 to $12 per million output tokens. Claude Sonnet 5, GPT-5.6 Terra and Gemini 3.1 Pro Preview all sit there. For teams building agents, that convergence makes list price the least useful number on the bill: cost per completed task now depends more on tokenizers, caching, output length and retry rates than on which vendor you pick.

If you are choosing a model for an agent by comparing rate cards, you are optimizing a variable that has stopped moving. The variables that still move (tokens burned per task, share cached, and how often runs fail) only come from measuring your own workload. Below are the verified September 2026 prices, two worked cost-per-task scenarios, and a checklist for keeping an agent budget honest.

What do frontier model APIs cost as of September 2026?

Three events shaped the current price map. First, Anthropic launched Claude Sonnet 5 on June 30, 2026 at what it called introductory pricing of $2 input and $10 output per million tokens, with a scheduled rise to $3/$15 on September 1. That rise never happened: Anthropic's pricing documentation now states that the $2/$10 price "is now the standard price" and that "the previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur." Anthropic's Sonnet 5 launch post carries the same update.

Second, OpenAI moved to a tiered naming scheme with GPT-5.6. Its developer documentation describes Sol as a "flagship model for complex professional work," recommends Terra "to balance intelligence and cost," and positions Luna "for cost-sensitive, high-volume workloads." Its API pricing page lists Terra at the same $2 input as Sonnet 5, and notes that "GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026."

Third, DeepSeek kept pushing the bottom of the market down. Its change log records the release of DeepSeek-V4.1-Flash on September 10, 2026, with the note that "API prices have been reduced accordingly," after introducing peak and off-peak pricing in August.

Here is the standard, pay-as-you-go rate card, taken from each vendor's own pricing page as of September 2026. All figures are US dollars per million tokens.

Two patterns stand out. The workhorse tier has collapsed onto a single input price, with output differing by only $2. And the gap between the workhorse and flagship tiers is now 2x to 2.5x on list price, down from the 5x spread between Sonnet-class and Opus-class models that Anthropic's older Opus 4.1 ($15/$75) still shows on the same page.

Why is the price floor converging around $2 per million input tokens?

The vendors do not explain their pricing decisions, so what follows is analysis rather than reported fact. The most plausible reading is that the workhorse tier has become the default for agent workloads, and agent workloads are where buyers compare prices most carefully. An agent that runs thousands of multi-step tasks a day turns a $1 difference in input price into a line item a finance team notices. Anthropic's own documentation now tells developers to "choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning," which is a direct acknowledgement of where the volume sits.

Anthropic reversing a pre-announced price increase is the clearest signal. A vendor that planned a 50% rise and then cancelled it was, at minimum, not confident the higher price would hold against GPT-5.6 Terra and Gemini 3.1 Pro at $2. The floor is being set by competition in that tier, while the budget tier (DeepSeek, Luna, Flash models) keeps pulling expectations down from below. We covered the open-weight side of that pressure in our look at the open-source LLM landscape.

Why is price per token the wrong number for agents?

Because an agent does not buy tokens. It buys completed tasks, and four factors separate the rate card from the invoice.

1. Tokenizers are not equal

The same text becomes a different number of tokens on different models. Anthropic's pricing page says Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text," and its Sonnet 5 launch post puts the range at roughly 1.0x to 1.35x, depending on the content type. A 35% swing in token count is larger than the 20% output-price gap between Sonnet 5 and Terra ($10 vs $12). The only way to compare is to run your real prompts through each vendor's token counter.

2. Caching changes the math more than list price

Agent loops resend the same system prompt, tool definitions and conversation history on every turn. Cached reads are priced at a steep discount everywhere: 0.1x input on most Claude models and on OpenAI's GPT-5.6 tiers, 0.025x on Claude Fable 5.1, and on DeepSeek's Flash model a cache hit costs $0.006 against $0.30 for a miss at peak. Anthropic charges extra to write the cache (1.25x input for a five-minute cache, 2x for one hour), and OpenAI's pricing page now lists a separate cache-write price as well. Google charges an hourly storage fee for cached context on Gemini 3.1 Pro.

3. Output dominates long runs

Output tokens cost four to six times input on every model in the table. Agents that reason out loud, write long tool calls or produce code files spend most of their budget on output. A model that finishes in fewer, shorter steps can beat a cheaper model that rambles.

4. Long context and data residency carry surcharges

OpenAI charges a higher "long context" rate on its GPT-5.6 and GPT-6 models (double the input price on Sol and Terra), and Google charges more for Gemini 3.1 Pro prompts above 200,000 tokens. Anthropic's pricing page says its 4.6-and-later models include the full 1M-token window at standard pricing. Both Anthropic and OpenAI add 10% for US-only or regional processing on newer models.

What does one agent task actually cost on each model?

To make the trade-offs concrete, here are two illustrative workloads. These are our own arithmetic on published list prices, not benchmarks, and they assume identical token counts on every model, which the tokenizer point above tells you will not be true in practice.

• Scenario A (single pass): 100,000 uncached input tokens and 10,000 output tokens. Think of a document review or one large code-generation call.
• Scenario B (20-turn agent loop): each turn reads a 50,000-token cached prefix, adds 5,000 fresh input tokens and writes 2,000 output tokens. Totals: 1M cached tokens, 100,000 fresh input, 40,000 output. This excludes the one-time cache-write cost and Gemini's cache storage fee.

Three readings from the table:

• At the workhorse tier, the vendors are within about 10% of each other. Sonnet 5, Terra and Gemini 3.1 Pro land at $0.30 to $0.32 for Scenario A and $0.80 to $0.88 for Scenario B. That gap is smaller than the tokenizer variance, so it should not decide your choice.
• Caching reorders the top tier. Fable 5.1 and GPT-6 Astra cost the same in Scenario A, but Fable 5.1's cheaper cache reads make it $0.75 less in the cached loop.
• The budget tier is an order of magnitude cheaper. Luna and DeepSeek Flash run Scenario B for under a dime. If a budget model can do a step reliably, routing that step down is the single biggest saving available.

How should retries and success rates change the comparison?

The cost that matters is cost per successful task: cost per attempt divided by the rate at which attempts succeed. A hypothetical example shows why. A $0.30 attempt that succeeds 60% of the time costs $0.50 per success. A $0.75 attempt that succeeds 90% of the time costs about $0.83. The cheaper model still wins in that case, but if the cheaper model's failures are silent (a wrong answer the agent does not notice), the downstream cost of catching them can erase the gap.

Vendors publish capability benchmarks, not success rates on your task. Anthropic put Sonnet 5 at 63.2% on an agentic coding evaluation, against 69.2% for Opus 4.8 and 58.1% for Sonnet 4.6, according to TechCrunch's launch coverage. Numbers like these tell you which way the gap points, not how big it is on your workload. For how to structure multi-model pipelines around those gaps, see our decision tree for agentic design patterns.

Which prices are still likely to move?

"Floor" does not mean "frozen." Several of the current prices come with an expiry date or a change already announced by the vendor:

• GPT-5.6 Sol: OpenAI labels its current rate promotional, available "at least through November 21, 2026." Budget for a possible increase after that date.
• Gemini 3.6, 3.7 and 3.8 Flash: Google's pricing page shows input and output prices doubling on January 1, 2027, from $0.75/$3.75 to $1.50/$7.50.
• Gemini 3.1 Pro: still labelled Preview, which means terms can change before general availability.
• DeepSeek: from 12:00 Beijing time on September 14, 2026, DeepSeek says requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at the Flash price until a V4.1 Pro ships. Peak hours, when the full rate applies, are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday.
• Claude Sonnet 5: the one price in this group that moved toward stability, from introductory to standard.

The same dynamic played out one layer up the stack in the AI coding tool pricing wars: introductory and promotional prices become the reference point buyers anchor to, and vendors find it hard to walk them back.

What should teams building agents do now?

1. Measure tokens per task on each candidate model. Run 50 to 100 representative tasks through each vendor, record input, cached, and output tokens plus success rate, and compute cost per successful task. Rate cards alone will mislead you by up to a third because of tokenizer differences.
2. Design for the cache. Put stable content (system prompt, tool schemas, reference documents) at the front of every request so it is cached. On a 20-turn loop, cached reads are the largest token category and the cheapest.
3. Route by step, not by agent. Use a workhorse model for planning and judgment, and push extraction, classification and formatting steps to a budget model where your evaluation shows it succeeds.
4. Batch anything that is not interactive. Anthropic, OpenAI and Google all list roughly 50% discounts for batch processing; OpenAI also offers a Flex tier at the same discount.
5. Put price-change dates in your calendar. September 14, 2026 (DeepSeek V4 Pro routing), November 21, 2026 (earliest end of Sol's promotional price) and January 1, 2027 (Gemini Flash increase) are all published. Rerun your cost model before each one.
6. Keep a second vendor wired in. With workhorse prices this close, an abstraction layer that can move a step between providers is cheap insurance against an unplanned price change.

The bottom line

The frontier API market in September 2026 has four tiers with clear price anchors: roughly $10/$50 at the top, $4 to $5 input for flagships, $2 input for workhorses, and cents for the budget tier. Within a tier, list-price differences are now smaller than the differences created by tokenizers, caching and retries. Choosing on the rate card is choosing on noise. Choose on measured cost per successful task, and re-measure whenever a promotional price expires.

Frequently Asked Questions

Is Claude Sonnet 5's $2/$10 pricing permanent?

Yes, as of September 2026. Anthropic launched Sonnet 5 at $2 per million input tokens and $10 per million output tokens as introductory pricing through August 31, 2026, with a planned rise to $3/$15. Its pricing documentation now says $2/$10 is the standard price and the September 1 increase will not occur. As with any API, Anthropic can still change prices in the future.

Which frontier model is cheapest for AI agents?

On list price, GPT-5.6 Luna and DeepSeek V4.1 Flash are the cheapest models in this comparison, both costing cents per task in our scenarios. Among workhorse models, Claude Sonnet 5, GPT-5.6 Terra and Gemini 3.1 Pro Preview are within about 10% of each other. The cheapest option for your agent depends on measured tokens per task and success rate, not the rate card.

How much does prompt caching save on agent workloads?

A lot, because agents resend the same context every turn. Cached reads cost 10% of the input price on most Claude models and on GPT-5.6, 2.5% on Claude Fable 5.1, and 2% on DeepSeek's Flash model. In our 20-turn scenario, 1 million of the 1.14 million total tokens were cached reads. Budget for cache-write charges and, on Gemini, hourly storage fees.

Why do the same prompts cost different amounts on different models?

Each vendor uses its own tokenizer, so identical text becomes a different number of billable tokens. Anthropic says its newer tokenizer, used by Claude 4.7 and later models, produces roughly 30% more tokens for the same text. Output length and the number of steps a model takes to finish also vary. Test with your real prompts and each vendor's token-counting tool before comparing costs.

When could these API prices change?

Several changes are already published. OpenAI says GPT-5.6 Sol's promotional pricing lasts at least through November 21, 2026. Google's Gemini 3.6, 3.7 and 3.8 Flash prices double on January 1, 2027. DeepSeek routes V4 Pro requests to V4.1 Flash pricing from September 14, 2026. Recheck vendor pricing pages before each date and rerun your cost model.