Open vs closed AI models: what they actually cost in 2026
In July the open tier cost five to fifty times less than the closed frontier. By October the gap within each price tier was four times at most, because closed labs cut prices while the biggest open models got more expensive. One open flagship now costs more per session than the closed workhorses. Here are the current rates, tier by tier, and what changes once caching is in the math.
In July we wrote that the strongest open-weight models cost five to fifty times less than the closed frontier. That was true then. Three months later it is not, at least not in the way most people mean it.
Two things moved at once. The closed labs cut prices hard: OpenAI and Anthropic now sell their main day-to-day models, GPT-6.1 Sol and Claude Sonnet 5.5, at $2 input and $10 output per million tokens. In July the comparable models, GPT-5.5 and Claude Sonnet 4.6, charged $5 / $30 and $3 / $15. Meanwhile the newest open flagships got much bigger and priced accordingly. Moonshot's Kimi K3 lists at $15 per million output tokens, more than either GPT-6.1 Sol or Claude Sonnet 5.5.
The result is a market where "open" no longer reliably means "cheap." Compared tier for tier, open models still usually win on price, but by four times at most rather than five to fifty. This piece lays out the current per-token rates on both sides as of October 11, 2026, compares them like for like, and works through what an agent session actually costs once prompt caching is in the math.
What changed since July
- Closed mid-tier prices fell by a third to two thirds. GPT-6.1 Sol and Claude Sonnet 5.5 both list at $2 / $10. In July the comparable models were GPT-5.5 at $5 / $30 and Claude Sonnet 4.6 at $3 / $15.
- Closed small models now price like open flash models. GPT-6 Luna and Claude Haiku 5.5 both list at $0.10 / $0.50.
- The biggest open models got more expensive. Kimi K3 lists at $3 / $15, against $0.95 / $4 for Kimi K2.6. GLM 5.3 lists at $1.40 / $4.40, against $1.00 / $3.20 for GLM 5.
- DeepSeek's cheapest model costs more. DeepSeek V4 Flash ($0.14 / $0.28) has been retired. Its replacement, V4.1 Flash, costs $0.15 to $0.30 input and $0.60 to $1.20 output depending on time of day.
- The top of the closed market moved up. Claude Fable 5.1 and GPT-6 Astra both list at $10 / $50, well above the $30 output ceiling of our July table.
- Introductory pricing is spreading. Google's Gemini 4 Argon launched at a rate that doubles on December 1, 2026, Gemini 3.8 Flash doubles on January 1, 2027, and OpenAI's GPT-5.6 Sol is on a promotional rate.
How model pricing actually works
Model APIs bill per token, and almost always at two different rates: one for input tokens (the prompt you send, including documents, system instructions, and prior conversation) and a higher one for output tokens (what the model generates). A token is roughly three-quarters of a word. Prices are quoted per million tokens.
The split matters because the ratio of input to output in your workload decides which price dominates your bill. A summarization job that reads a long document and writes a short answer is input-heavy. An agent that writes a lot of code or analysis is output-heavy.
Most providers also charge a much lower cached input rate for prompt content they can reuse from a previous call. This matters more than it used to. An agent re-sends its whole conversation on every turn, so in a long session the large majority of input tokens are cache hits. For that kind of work, the cached rate can matter as much as the headline input price.
Three newer wrinkles also show up on the rate cards: long-context surcharges (a higher rate once a prompt passes a size threshold), time-of-day pricing (DeepSeek charges double during peak hours), and introductory prices that expire. We cover each below.
The closed models
These are the proprietary models from the four largest closed labs, at each lab's published standard rate per million tokens, for prompts under any long-context threshold.
| Model | Lab | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10.00 | $50.00 | $0.25 |
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | $1.00 |
| Claude Opus 5.5 | Anthropic | $4.00 | $20.00 | $0.20 |
| Gemini 3.1 Pro Preview | $2.00 | $12.00 | $0.20 | |
| Gemini 4 Argon | $2.00 | $10.00 | $0.10 | |
| GPT-6.1 Sol | OpenAI | $2.00 | $10.00 | $0.10 |
| Claude Sonnet 5.5 | Anthropic | $2.00 | $10.00 | $0.10 |
| Grok 4.7 | xAI | $2.00 | $6.00 | $0.50 |
| Gemini 3.8 Flash | $0.75 | $3.75 | $0.075 | |
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | $0.01 |
| Claude Haiku 5.5 | Anthropic | $0.10 | $0.50 | $0.01 |
Gemini 3.8 Flash's rate is introductory through December 31, 2026, and rises to $1.50 / $7.50 on January 1, 2027. Claude Haiku 5.5's rate applies to prompts up to 100,000 tokens. Gemini 4 Argon became generally available on October 13, 2026. Its rate is introductory through November 30, 2026, and rises to $4 / $20, with cached input at $0.20, on December 1, the same as Claude Opus 5.5.
The closed market now has four clear price points: a premium tier at $50 output (Fable 5.1 and Astra), a flagship at $20 (Opus 5.5), a dense band of workhorses at $6 to $12, and small models at $0.50.
The open-weight models
These models publish their weights, so in principle anyone can host them. Rates below are each lab's own first-party API price unless noted. Several of the strongest still come from Chinese labs: DeepSeek, Alibaba (Qwen), Zhipu / Z.ai (GLM), Moonshot (Kimi), and MiniMax. The maintained list of what Perch runs is on the Models page.
| Model | Lab | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|---|
| Kimi K3 | Moonshot | $3.00 | $15.00 | $0.30 |
| GLM 5.3 | Zhipu / Z.ai | $1.40 | $4.40 | $0.26 |
| Inkling | Thinking Machines Lab | $1.00 | $4.05 | $0.17 |
| Kimi K2.7 Code | Moonshot | $0.95 | $4.00 | $0.19 |
| DeepSeek V4 Pro (peak) | DeepSeek | $1.32 | $3.96 | $0.044 |
| Qwen 3.8 27B | Alibaba | $0.50 | $3.00 | about $0.10 |
| MiniMax M3 | MiniMax | $0.30 | $1.20 | $0.06 |
| DeepSeek V4.1 Flash (peak) | DeepSeek | $0.30 | $1.20 | $0.006 |
| GLM 5.3 Flash | Zhipu / Z.ai | $0.15 | $0.50 | $0.03 |
| Laguna S 2.1 | Poolside | $0.09 | $0.18 | $0.009 |
DeepSeek's off-peak rates are half the peak rates shown. Inkling's price is for Thinking Machines' Tinker serverless inference, which is in beta. Poolside does not publish a price page; the Laguna S 2.1 rate is Poolside's own endpoint as listed on OpenRouter.
"Open" has also become a fuzzier label. Alibaba's hosted Qwen 3.8 Max 0902 ($2 / $6) and Qwen 3.8 Flash are API-only builds, even though open base models sit underneath them. Kimi K3's weights are public, but its license requires a separate agreement for model-as-a-service businesses above $20 million in revenue. GLM 5.3 ships under a custom license rather than MIT. Check the license, not just the download link, if self-hosting is part of your plan.
Compare by tier, not by open versus closed
Lining up the cheapest open model against the most expensive closed one produces a dramatic number, and in October 2026 that number is bigger than it was in July: $0.18 against $50 per million output tokens, a factor of about 280. It is also not a useful comparison, because nobody is choosing between those two models for the same job.
Compared within a price tier, the picture is much closer:
| Tier | Closed, output / 1M | Open, output / 1M | Gap |
|---|---|---|---|
| Flagships | $20 to $50 | $15 | 1.3x to 3.3x |
| Workhorses | $6 to $12 | $3.00 to $4.40 | 1.4x to 4x |
| Low-cost | $0.50 | $0.18 to $1.20 | roughly even |
The flagships are Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra against Kimi K3. The workhorses are Grok 4.7, GPT-6.1 Sol, Claude Sonnet 5.5, Gemini 3.1 Pro, and Gemini 4 Argon at its introductory rate against Qwen 3.8 27B, DeepSeek V4 Pro at peak, Kimi K2.7 Code, Inkling, and GLM 5.3. The low-cost tier is GPT-6 Luna and Claude Haiku 5.5 against Laguna S 2.1, GLM 5.3 Flash, DeepSeek V4.1 Flash, and MiniMax M3.
These tiers group models by price, not by a claim that they are equally capable. Within a tier, quality still varies, sometimes a lot. But a tier is roughly the set of models a buyer weighs against each other, and inside each one the open discount is now a few times at most.
What an agent session actually costs
Per-token rates do not tell you what a job costs. To make the numbers concrete, here is one illustrative agent session priced on each model: 1 million input tokens and 40,000 output tokens, spread over about 25 turns of 40,000 tokens each, which keeps every request under all of the long-context thresholds. We price it two ways: with 90 percent of input tokens served from cache, which is typical for an agent that re-sends its conversation every turn, and with no caching at all.
| Model | Weights | With caching | No caching |
|---|---|---|---|
| GPT-6 Astra | Closed | $3.90 | $12.00 |
| Claude Fable 5.1 | Closed | $3.23 | $12.00 |
| Claude Opus 5.5 | Closed | $1.38 | $4.80 |
| Kimi K3 | Open | $1.17 | $3.60 |
| Grok 4.7 | Closed | $0.89 | $2.24 |
| Gemini 3.1 Pro Preview | Closed | $0.86 | $2.48 |
| Gemini 4 Argon (introductory) | Closed | $0.69 | $2.40 |
| GPT-6.1 Sol | Closed | $0.69 | $2.40 |
| Claude Sonnet 5.5 | Closed | $0.69 | $2.40 |
| GLM 5.3 | Open | $0.55 | $1.58 |
| Kimi K2.7 Code | Open | $0.43 | $1.11 |
| DeepSeek V4 Pro (peak) | Open | $0.33 | $1.48 |
| Qwen 3.8 27B | Open | $0.26 | $0.62 |
| MiniMax M3 | Open | $0.13 | $0.35 |
| DeepSeek V4.1 Flash (peak) | Open | $0.08 | $0.35 |
| GLM 5.3 Flash | Open | $0.06 | $0.17 |
| GPT-6 Luna | Closed | $0.04 | $0.12 |
| Claude Haiku 5.5 | Closed | $0.04 | $0.12 |
| Laguna S 2.1 | Open | $0.02 | $0.10 |
Uncached input is billed at the standard input rate. Some providers charge a premium to write content into the cache, which would add a few cents to the cached column; we leave it out to keep the comparison simple.
Four things stand out:
- Caching compresses the open discount among workhorses. With caching, GPT-6.1 Sol and Claude Sonnet 5.5 cost $0.69 and the open workhorses $0.26 to $0.55, a gap of 1.25 to 2.7 times. Anthropic and OpenAI bill cache hits at 5 percent of the input price; GLM 5.3 bills about 19 percent.
- The largest open model costs more than the closed workhorses. Kimi K3 comes in at $1.17 with caching and $3.60 without, more than every closed workhorse in the table either way.
- A low output price can hide a high session cost. Grok 4.7 has the lowest output price of the closed workhorses, but bills cache hits at 25 percent of input, which makes it the most expensive of that group once caching is counted.
- The closed small models are among the cheapest options. GPT-6 Luna and Claude Haiku 5.5 cost about four cents for the cached session, less than DeepSeek V4.1 Flash at peak or GLM 5.3 Flash. Only Laguna S 2.1, and DeepSeek V4.1 Flash off-peak at about the same four cents, match or beat them.
Your own numbers will differ with your cache hit rate and your input-to-output mix. The method is the point: price the session, not the token.
Why the gap narrowed
Closed labs cut prices. OpenAI's changelog records a run of reductions across the summer, including an 80 percent cut on GPT-5.6 Luna in July and a promotional cut to $4 / $20 on GPT-5.6 Sol in August, before GPT-6 Sol and GPT-6.1 Sol launched at $2 / $10. Anthropic's current models are priced below their July counterparts at every tier: Opus 5.5 at $4 / $20 against Opus 4.8's $5 / $25, Sonnet 5.5 at $2 / $10 against Sonnet 4.6's $3 / $15, and Haiku 5.5 at $0.10 / $0.50 against Haiku 4.5's $1 / $5.
Open flagships got much bigger. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token, and the open base under Qwen 3.8 Max is 2.4 trillion. Serving cost tracks the active parameters, and a larger model costs more to run per token no matter who hosts it. The prices of the newest open flagships reflect that.
Licenses can limit the hosting competition. Part of what kept open models cheap was that many providers could host the same weights and compete on price. That still happens for permissively licensed models. But when a license restricts commercial hosting, as Kimi K3's does above a revenue threshold, fewer hosts can compete, and the lab's own price matters more.
None of this means the closed labs are now cheap. Their premium tier at $50 output is the most expensive in the market. It means the middle of the market, where most production work runs, has converged.
The fine print that changes the bill
The rate cards have grown more conditions. Five matter for budgeting:
- Long-context surcharges. OpenAI's GPT-6 family charges more for prompts above 272,000 input tokens (double on input, 1.5 times on output). Google's Gemini 3.1 Pro does the same above 200,000. xAI charges double on both sides from 200,000 tokens, applied to the whole request. Claude Haiku 5.5 charges five times more above 100,000. Claude Opus 5.5 and Sonnet 5.5 bill their full 1-million-token context at the standard rate. Google lists no long-context surcharge for Gemini 4 Argon, whose 1-million-token limit is shared between input and output.
- Time-of-day pricing. DeepSeek charges its peak rate from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and half that the rest of the time. A batch job scheduled off-peak costs half as much.
- Introductory and promotional prices. Gemini 4 Argon doubles on December 1, 2026, and Gemini 3.8 Flash on January 1, 2027. GPT-5.6 Sol is on a promotional rate at least through November 21, 2026. MiniMax M3's rate is shown as a permanent 50 percent discount from its list price. Budget with the price you will pay next quarter, not this one.
- Cache-hit rates vary far more than base rates. As a share of the input price, cache hits cost 2 percent on DeepSeek V4.1 Flash, 2.5 percent on Claude Fable 5.1, 5 percent on Claude Opus 5.5, Sonnet 5.5, GPT-6.1 Sol, and Gemini 4 Argon, 10 percent on Kimi K3 and GPT-6 Astra, about 19 percent on GLM 5.3, and 25 percent on Grok 4.7. For agent work, compare this column before the headline one.
- Reasoning modes multiply tokens, not rates. OpenAI's "pro" reasoning mode bills at the model's standard rate but generates more tokens. A model that thinks longer costs more per task at the same per-token price.
Where each side still earns its place
The low-cost open models are still the cheapest way to run volume. For high-throughput, latency-tolerant work like extraction, classification, and first-pass review, Laguna S 2.1, GLM 5.3 Flash, and DeepSeek V4.1 Flash off-peak are hard to beat on price, and open weights let you self-host or fine-tune when the data cannot leave your environment.
The open workhorses are still a real discount. Up to four times cheaper per output token, and 1.25 to 2.7 times cheaper per cached session than GPT-6.1 Sol or Sonnet 5.5 in our example, is meaningful at volume. In our July coding benchmark, an open-weight model won one of two tasks outright, so the discount does not automatically mean a quality sacrifice. Test on your own work.
The closed small models deserve a second look. If you last priced them when Claude Haiku 4.5 cost $5 per million output tokens, the math has changed. At $0.50 output with cheap caching, they now compete directly with the open flash tier.
The hardest work still favors the top closed tier. On the most demanding reasoning and long-horizon agent tasks, the premium closed models remain the most reliable choice, and if a task genuinely needs them, paying for them is cheaper than paying for retries.
Cost per finished task is the only number that settles it. A cheaper model that needs two attempts, or that writes more tokens to reach the same answer, can erase its per-token advantage. In our forensics comparison, a model that was 4.2 times more token-efficient than Opus 4.8 on a coding benchmark cost twice as much as Opus 4.8 on document-heavy financial work. Benchmarks measure the workload they were run on.
How we think about it
Perch does not pick a side in the open-versus-closed debate, because it is the wrong frame, and it is a less useful one every quarter. The useful question is which model fits the task in front of you, at what cost per finished task. So Perch is model-agnostic: Roost routes each task to a capable model automatically, and you can pin a specific one by hand. The full lineup is on the Models page with the rate we pay our inference provider for each model, which is the rate we meter usage at, with no markup. For a few models that rate differs from the lab list prices in this article; Gemini 4 Argon, for example, is metered at its $4 / $20 standard rate rather than the introductory one. If you already pay a provider directly, you can bring your own key on any plan.
That is also why we can run comparisons like this one honestly. We are not defending a house model. We evaluate every candidate the same way, on realistic tasks with known-correct answers, and let the price sit next to the result. In October 2026, the result is that open and closed models overlap across most of the price range, and the right choice depends on the job.
For plans and what is included on each, see Pricing.
Method and sources
Prices are each lab's published standard API rates in US dollars per million tokens, checked on October 11, 2026, for prompts under any long-context threshold. Where a lab publishes no per-token price, we say so and name the source we used. Gemini 4 Argon, which became generally available afterwards, was added on October 13, 2026. The July 2026 figures quoted for comparison are from the original version of this article. The worked session example is illustrative; its assumptions are stated above the table.
- OpenAI: API pricing and changelog
- Anthropic: Claude API pricing
- Google: Gemini API pricing
- xAI: Models and pricing
- DeepSeek: API pricing
- Alibaba: Model Studio pricing
- Zhipu / Z.ai: API pricing
- Moonshot: Kimi API pricing
- MiniMax: Pay-as-you-go pricing
- Thinking Machines Lab: Tinker models
- Poolside: Laguna S 2.1 on OpenRouter
Frequently asked questions
- Are open-source AI models cheaper than GPT and Claude?
- Usually, but by much less than earlier in 2026. Compared tier for tier in October 2026, open-weight flagships and workhorses cost roughly 1.3 to 4 times less per output token than their closed counterparts. Closed workhorses like GPT-6.1 Sol and Claude Sonnet 5.5 charge $10 per million output tokens, while open workhorses like GLM 5.3, Kimi K2.7 Code, and Qwen 3.8 27B charge $3 to $4.40. At the low-cost end, closed models like GPT-6 Luna and Claude Haiku 5.5 ($0.50 output) are now priced in the same range as open models like DeepSeek V4.1 Flash and GLM 5.3 Flash. And the largest open model, Kimi K3, costs more per output token than either GPT-6.1 Sol or Claude Sonnet 5.5.
- Are Chinese open models like DeepSeek, Qwen, and Kimi still cheaper?
- Mostly, but not across the board. GLM 5.3 ($1.40 input, $4.40 output per million tokens), Qwen 3.8 27B ($0.50 and $3.00), and DeepSeek V4 Pro ($1.32 and $3.96 at peak, half that off-peak) all undercut the closed workhorses. But Moonshot's Kimi K3 lists at $3 input and $15 output, above GPT-6.1 Sol and Claude Sonnet 5.5 at $2 and $10. And DeepSeek's cheapest model now costs $0.60 to $1.20 per million output tokens depending on time of day, up from $0.28 for the July model it replaced.
- How much cheaper are open models than closed models?
- On list output prices, about 1.3 to 3.3 times cheaper at the flagship tier and 1.4 to 4 times at the workhorse tier, with the low-cost tier roughly even. On a cache-heavy agent session, our worked example puts GPT-6.1 Sol and Claude Sonnet 5.5 at $0.69 against $0.26 to $0.55 for the open workhorses, a gap of 1.25 to 2.7 times. That is a long way from the five to fifty times gap of mid-2026.
- What is the cheapest capable AI model API in October 2026?
- At the low-cost end, Poolside's Laguna S 2.1 lists at $0.09 input and $0.18 output per million tokens on Poolside's own endpoint, GLM 5.3 Flash at $0.15 and $0.50, and DeepSeek V4.1 Flash at $0.15 and $0.60 off-peak. On the closed side, GPT-6 Luna and Claude Haiku 5.5 both list at $0.10 input and $0.50 output, and with prompt caching they are among the cheapest per session of any model we priced. Which is cheapest for you depends on your workload's mix of fresh input, cached input, and output.
- Why did the price gap between open and closed AI models shrink?
- Three reasons. Closed labs cut prices sharply. GPT-6.1 Sol and Claude Sonnet 5.5 both list at $2 input and $10 output, where the comparable models in mid-2026, GPT-5.5 and Claude Sonnet 4.6, charged $5 / $30 and $3 / $15. The newest open flagships are far larger, with Kimi K3 at 2.8 trillion total parameters, and cost more to serve. And some open licenses now restrict commercial hosting, which can limit the price competition between hosts that used to push open-model prices down.
- Does prompt caching change which model is cheapest?
- Yes, often more than the headline rate does. Agent workloads re-send the same context every turn, so most input tokens are cache hits. Claude Sonnet 5.5 and GPT-6.1 Sol bill cache hits at 5 percent of their input price, while GLM 5.3 bills about 19 percent and Grok 4.7 25 percent. In our worked example, Grok 4.7 has the lowest output price of the closed workhorses but the highest session cost once caching is included.