The Cheapest Near-Frontier AI Models (October 2026), Ranked by Cost per Task

Updated 2026-10-26: We now rank models by Artificial Analysis's cost per Intelligence Index task (v4.3.2) instead of per-token sticker price. The August line-up (DeepSeek V4-Flash, GPT-5.6 Luna, Qwen 3.7 Flash, Claude Opus 5) is replaced by October's models, including GPT-6.1 Sol, Claude Opus 5.5, Gemini 4 Argon, Kimi K3.1 Pro, Muse Wonder and DeepSeek V4.1 Flash 1017. We also cover Gemini 4 Argon's price doubling on December 1.

A month of releases rebuilt the top of the market. Anthropic shipped Claude Opus 5.5 on September 22. Google shipped Gemini 4 Argon on September 30, a day after OpenAI's GPT-6.1 Sol. Meta's Muse Wonder followed on October 12, DeepSeek's V4.1 Flash 1017 on October 17 and Moonshot's Kimi K3.1 Pro on October 23 (release dates from Artificial Analysis model pages). Almost all of these are priced between $2 and $4 per million input tokens. That makes the sticker price a poor guide to what a finished task will cost you.

Price per token doesn't show three things: how many tokens the model spends thinking, how many attempts it needs, and how far the effort setting pushes both up. This page ranks frontier-quality models by cost per task, using Artificial Analysis (AA), an independent benchmarking firm that runs every model through the same Intelligence Index and publishes what each task cost. If you want small, fast models ranked by token price and throughput, see our separate guide to the cheapest fast LLM APIs. This page covers the expensive end of the market and looks for the best value inside it.

What counts as "near-frontier" in October 2026?

We use a simple, checkable definition. A model is near-frontier if it scores within six points of the top of the Artificial Analysis Intelligence Index v4.3.2. As of October 26 the top is Claude Opus 5.5 at max effort with 58, so the bar is 52 or higher. DeepSeek V4.1 Flash 1017 scores 45, which puts it below the bar. We include it anyway as a budget anchor, because many teams will pick it over every model above it.

The index is a weighted average of 10 evaluations in four groups, according to AA's methodology page:

• Agents (30%): AA-Briefcase, GDPval-AA, AutomationBench-AA.
• Coding (20%): Terminal-Bench 4.0, SciCode.
• General (30%): AA-Omniscience, GDP.pdf, AA-LCR.
• Scientific reasoning (20%): Humanity's Last Exam, CritPt.

The index measures broad capability, with a heavy agentic and coding weighting. That fits most engineering workloads, but it won't predict your exact task. Use it to build a shortlist, then make the final call with your own evals.

The ranking: cost per Intelligence Index task

All index scores, cost-per-task figures, output-token counts and response times come from Artificial Analysis, as of October 2026. They are independent measurements, not vendor claims, and were last checked on October 26, 2026. AA updates them continuously: during October its cost-per-task figures for Gemini 4 Argon and Muse Wonder were revised, so re-check the live leaderboard before you budget on any single number. List prices are per 1M tokens and come from vendor pricing pages, except Kimi K3.1 Pro and Muse Wonder, which AA reports.

"n/a" means AA's model page lists output tokens only for the max configuration of Opus 5.5. "Total response time" is AA's measured end-to-end time on its leaderboard. It is a live median that moves daily: GPT-6.1 Sol max went from about 255 s to 333 s in the days we checked it. Treat it as a relative latency signal, not your production latency.

Which models are on the efficient frontier?

Plot score against cost per task and a clear frontier appears. On that frontier, every step up in score costs more, and every model off it is beaten by something cheaper that scores at least as well:

• GPT-6.1 Sol covers the low end at every effort level: low (42, $0.13), medium (48, $0.21), high (50, $0.32), xhigh (51, $0.39), max (52, $0.72).
• Gemini 4 Argon high (53, $1.62), until December 1.
• Claude Opus 5.5 high (54, $1.82).
• Kimi K3.1 Pro (55, $2.11).
• Claude Opus 5.5 xhigh (56, $3.46).
• Claude Opus 5.5 max (58, $5.98).

The notable model off the frontier is Muse Wonder (52, $8.63). It matches GPT-6.1 Sol max on score at about 12x the cost per task ($8.63 / $0.72), and it costs more per task than Opus 5.5 max, which scores 6 points higher. It can still be the right choice for reasons this index doesn't capture: Muse Wonder has the highest output speed AA reports among these models (about 197 tokens/s), though Opus 5.5 high finishes sooner end to end (40.1 s vs 49.3 s). Gemini 4 Argon is on the frontier today, but only just, and its December 1 price doubling would push it off (see below).

Model by model: where each one earns its place

1. GPT-6.1 Sol (max): the value pick for near-frontier work

GPT-6.1 Sol lists at $2 input, $0.10 cached input and $10 output per 1M tokens for prompts up to 272K tokens. Above 272K it's $4 / $0.20 / $15. It is the cheapest model to clear the near-frontier bar, and the reason is efficiency, not a discount. AA measured 67M output tokens across the index, below the 81M median for its class. Gemini 4 Argon and Kimi K3.1 Pro list at the same $2 / $10, yet Sol's cost per task is 44% of Argon's ($0.72 / $1.62) and about a third of Kimi's ($0.72 / $2.11 = 0.34).

The trade-off is latency at max effort. AA recorded about 326 seconds to first chunk and 333 seconds end to end, more than five minutes. That's fine for batch jobs, code review or overnight agents, but too slow for anything interactive. Dropping to high effort costs you 2 index points (52 to 50) and cuts the cost per task to $0.32 and the total response time to about 83 seconds. Full details are in our GPT-6.1 Sol guide.

2. Gemini 4 Argon: the cheapest route to 53, until December 1

Gemini 4 Argon (gemini-4-argon) scores 53 for $1.62 per task, which makes it the cheapest way to reach 53 on AA's index today and puts it on the efficient frontier between GPT-6.1 Sol max and Opus 5.5 high. AA counted 110M output tokens, which it describes as somewhat verbose against an 81M median. Google's pricing page lists Standard-tier rates of $2 input, $10 output and $0.10 for implicit context caching per 1M tokens through November 30, 2026. On December 1, every price doubles across all tiers:

Every price category doubles, so if token usage stays the same, cost per task also roughly doubles. That's about $3.24 per task ($1.62 × 2). This is our estimate, not an AA measurement. At that level Argon (53) would cost more per task than both Opus 5.5 high (54, $1.82) and Kimi K3.1 Pro (55, $2.11), which score higher, and its list price would match Opus 5.5 exactly. AA's own Argon figure also moved during October (an earlier snapshot we used showed $1.99), another reason to re-check before you budget. Argon has no free tier and supports only implicit caching. If Argon is in your stack for cost reasons, re-run the numbers before December. Our Gemini 4 Argon guide covers the rest.

3. Claude Opus 5.5 (high): the cheapest route to 54

Opus 5.5 lists at $4 input and $20 output, with cache hits at $0.20 per 1M. That's 5% of base input (0.05x), versus the standard 0.1x on most other Claude models, according to Anthropic's pricing page. The full 1M-token context is billed at standard rates. At high effort, AA measured 54 points for $1.82 per task, with a 40.1-second total response. That makes it the fastest near-frontier configuration in the table.

The effort setting is the real pricing control on Opus 5.5:

Going from high to max adds 4 points and multiplies cost per task by 3.3 ($5.98 / $1.82). The extra spend is almost all output: AA counted 260M output tokens for Opus 5.5 max, about 3.2 times the 81M median. A good default is to run high and send only the hardest tickets to xhigh or max. See our Claude Opus 5.5 guide for how to set effort per request.

Claude Sonnet 5.5 ($2 / $10, cache hits $0.10) is worth knowing about. AA measured 52 points for $2.01 per task at xhigh and 56 for $5.46 at max. Both sit behind Opus 5.5 at similar spend, so on this index Opus is the better buy per point.

4. Kimi K3.1 Pro: the highest score under $2.50 per task

Kimi K3.1 Pro (released October 23) scores 55, one point above Opus 5.5 high, for $2.11 per task. AA lists it at $2 input / $10 output with a 1M-token context, as a proprietary model with no public weights.

Kimi K3.1 Pro shows clearly how verbosity drives cost. It spent about 150M output tokens on the index, against an 81M median. AA also measured about 50 tokens/s, against a median of about 83 for its price tier. Its list price matches GPT-6.1 Sol, but it produces about 2.2x as many output tokens (150M / 67M) and costs 2.9x as much per task ($2.11 / $0.72). It is also slow to stream long answers. If you need the extra 3 points over Sol max, it's the cheapest way to get them. Read our Kimi K3.1 Pro guide before you commit.

5. Muse Wonder: fast, but expensive per point

Meta's Muse Wonder (max) scores 52 at $8.63 per task. AA lists it at $4 / $20 with a 95% cache discount and measured 130M output tokens. Its strength is output speed: about 197 tokens/s, the highest of the models here, with a 49.3-second total response on AA's test. Pick it when streaming speed matters more than cost per task. If cost is the goal, GPT-6.1 Sol max reaches the same score for $0.72, and Opus 5.5 max scores 58 for less ($5.98). AA revised this cost figure during October (an earlier snapshot we used showed $4.52), so check the live page before you rely on it. AA lists a "Muse Wonder Full (max)" at 56 but publishes no cost data for it, so we leave it out of the ranking. See our Muse Wonder guide.

Budget anchor: DeepSeek V4.1 Flash 1017

DeepSeek V4.1 Flash 1017 scores 45 for $0.26 per task, under half of GPT-6.1 Sol max's $0.72. It is the only model on this page with open weights (MIT). DeepSeek serves it under the deepseek-flash model ID. Peak pricing is $0.30 input (cache miss), $0.006 cache hit and $1.20 output per 1M tokens. Off-peak rates are 50% of peak. DeepSeek defines peak as 01:00–04:00 and 06:00–10:00 UTC on weekdays (Chinese public holidays are off-peak), so batch work scheduled outside those hours costs less than the list price (AA's model page uses the $0.30 / $1.20 rates). It's also fast, at 169 tokens/s and a 16-second total response on AA's test.

DeepSeek is verbose too. AA counted 210M output tokens against a 140M median for open-weights models. The low per-token price absorbs that extra output, but it shows up again if you self-host and pay for GPU time.

A caveat: on this index, GPT-6.1 Sol at medium effort (48, $0.21) beats DeepSeek V4.1 Flash on both score and cost per task. DeepSeek still makes sense if you need open weights, self-hosting, low latency or off-peak batch pricing. If you only care about index points per dollar from a hosted API, Sol medium is the better value. See our DeepSeek V4.1 Flash guide for setup and self-hosting.

Why verbosity, not sticker price, drives cost per task

Three of the models above (GPT-6.1 Sol, Gemini 4 Argon, Kimi K3.1 Pro) have the same $2 / $10 list price. Their cost per task ranges from $0.72 to $2.11. The biggest difference is how many output tokens, including reasoning tokens, each model generates to finish a task:

Output tokens don't explain everything. Input volume, cache hit rates and the number of agent turns per task also matter, and AA folds measured cache hit rates into its cost figures. Muse Wonder is the clearest exception: it used half the output tokens of Opus 5.5 max at the same $20 output price, yet AA puts its cost per task higher ($8.63 vs $5.98), and the summary page doesn't break down why. For the three $2 / $10 models, though, the pattern holds. A model that thinks for twice as long at the same price costs about twice as much, and a 3x token gap swamps a 2x price gap.

Here's a simple illustration at a $10 per 1M output rate. One model finishes a task in 20,000 output tokens. A second model needs 2.24x as many (the Kimi-to-Sol ratio above), which is 44,800 tokens:

• Model A: 20,000 × $10 / 1,000,000 = $0.20
• Model B: 44,800 × $10 / 1,000,000 = $0.448

The per-token price is the same, but the output bill is 2.24x higher. Run your own tasks and log the real usage fields before you compare vendors on price sheets.

A token-mix calculator you can reuse

Sticker prices are still useful once you know your token mix. The example below assumes a typical agent step: 200K input tokens, of which 150K are a cached prefix, plus 20K output tokens, at October 2026 list prices. It ignores cache-write premiums (on Claude, a 5-minute cache write costs 1.25x base input), so treat it as the steady-state cost once the prefix is warm.

def task_cost(uncached_in, cached_in, out, price_in, price_cached, price_out):
    """All token counts are raw tokens; prices are USD per 1M tokens."""
    return (uncached_in * price_in + cached_in * price_cached + out * price_out) / 1_000_000

# 50K uncached input + 150K cached input + 20K output
mix = dict(uncached_in=50_000, cached_in=150_000, out=20_000)

prices = {
    "Claude Opus 5.5":            (4.00, 0.20, 20.00),
    "Claude Sonnet 5.5":          (2.00, 0.10, 10.00),
    "GPT-6.1 Sol (<=272K)":       (2.00, 0.10, 10.00),
    "Gemini 4 Argon (to Nov 30)": (2.00, 0.10, 10.00),
    "Gemini 4 Argon (Dec 1+)":    (4.00, 0.20, 20.00),
    "DeepSeek flash (peak)":      (0.30, 0.006, 1.20),
}

for name, (p_in, p_cached, p_out) in prices.items():
    print(f"{name:28s} ${task_cost(**mix, price_in=p_in, price_cached=p_cached, price_out=p_out):.4f}")

The arithmetic, line by line:

(Prices per 1M tokens; "50K × $4" means 50,000 × $4 / 1,000,000.) This is why the ranking uses cost per task. On a fixed token mix, Sol and Argon cost exactly the same today. On AA's real tasks, Argon costs 2.25x more ($1.62 / $0.72) because it spends more tokens getting to an answer. Prompt caching also matters more now that cache reads are 5% of input on Opus 5.5 and GPT-6.1 Sol. Put stable instructions, tool definitions and shared context at the front of the prompt so they stay cacheable.

What does a real monthly bill look like?

Take a team running 1,000 agentic tasks a month, each roughly as heavy as an average AA index task. Monthly cost is cost per task × 1,000:

Two things stand out. First, the spread is wider than per-token prices suggest. Opus 5.5 max costs 23x as much per task as DeepSeek, against a 13-17x gap in list prices, because it is pricier per token and also uses more tokens. Muse Wonder, at the same list price as Opus, comes out at 33x. Second, the gap inside the near-frontier band is large. Moving from GPT-6.1 Sol max to Opus 5.5 max adds 6 index points and $5,260 a month on this workload ($5,980 − $720). Whether those 6 points are worth it depends on what a failed task costs you. If a failure means an engineer spends an hour cleaning up, the expensive model can still come out ahead.

Your real tasks will differ from AA's. Short classification calls will cost far less than these figures, and long multi-hour agent runs can cost more. Use the ratios between models, not the dollar amounts.

When is the cheapest model not the right call?

• Interactive latency matters. GPT-6.1 Sol max has the best cost per point here, but AA measured more than five minutes to first chunk. For chat-style or in-IDE use, Opus 5.5 high (40.1 s) or Muse Wonder (49.3 s) will feel far better. Sol at medium (about 14 s) or high (about 83 s) effort is another option.
• The task is long, autonomous and expensive to fail. When a run spans dozens of tool calls with no human review, the extra index points from Opus 5.5 xhigh/max or Kimi K3.1 Pro can pay for themselves through fewer retries. Our Opus 5.5 vs GPT-6 Astra vs Grok 4.7 comparison covers the top tier in more depth.
• Data residency and compliance. DeepSeek's hosted API is a non-starter for many regulated teams. The MIT-licensed weights mean you can self-host V4.1 Flash instead. OpenAI charges a 10% uplift on regional-processing (data residency) endpoints for models released since March 5, 2026, which includes GPT-6.1 Sol, and Anthropic charges 1.1x for US-only inference via inference_geo, according to their pricing pages. Include that in your cost per task.
• Prices are about to move. Gemini 4 Argon doubles on December 1. Google's pricing page also shows Gemini 3.8 Flash and 3.6 Flash doubling on January 1, 2027. Re-check any cost model built on Google list prices before the end of the year.

How to pick

• Need near-frontier quality at the lowest cost, and latency isn't critical? GPT-6.1 Sol at max effort. Drop to high effort when you need faster responses.
• Need 54+ with reasonable latency? Claude Opus 5.5 at high effort, and route the hardest cases to xhigh.
• Need the highest score under about $2.50 per task? Kimi K3.1 Pro, as long as slow streaming is acceptable.
• Need the absolute best and can afford it? Opus 5.5 at max effort (58), applied only to tasks where failure is expensive.
• Need 53 for the least money, or already on Google Cloud? Gemini 4 Argon until November 30, but budget for the December 1 doubling.
• Need open weights, self-hosting or the lowest absolute bill? DeepSeek V4.1 Flash 1017, scheduled off-peak where you can.
• Want small, fast models instead? See the cheapest fast LLM APIs comparison.

Whichever you pick, build a 50-100 task eval from your own tickets, log token usage per task, and compute your own cost per successful task. That number should drive your model choice and your routing rules.

What this means for your engineering team

Cost per task only drops if someone builds the evals, routing rules and usage logging described above. That is engineering work, not a pricing decision: someone has to wire effort levels per request, keep prompts cache-friendly, and re-run the numbers when a vendor reprices, as Google does on December 1. Teams that treat model selection as an ongoing system rather than a one-time pick are the ones that capture the savings in this table.

FAQ

What is the cheapest near-frontier AI model in October 2026?

On Artificial Analysis's cost per Intelligence Index task, it's GPT-6.1 Sol at max effort: 52 points for $0.72 per task. It's the cheapest model within six points of the top score (58, held by Claude Opus 5.5 at max effort).

What is the cheapest model overall on this list?

DeepSeek V4.1 Flash 1017, at $0.26 per index task with a score of 45, served as deepseek-flash at $0.30 input / $1.20 output per 1M tokens at peak and half that off-peak. GPT-6.1 Sol at medium effort scores higher (48) for slightly less ($0.21 per task), so DeepSeek's main advantages are open weights, speed and off-peak pricing.

Why does Kimi K3.1 Pro cost more per task than GPT-6.1 Sol at the same token price?

Verbosity. Both list at $2 / $10 per 1M tokens, but AA measured about 150M output tokens for Kimi K3.1 Pro on the index versus 67M for GPT-6.1 Sol. More tokens per task means a higher bill: $2.11 vs $0.72 per task.

How much will Gemini 4 Argon cost after December 1, 2026?

Google's pricing page shows Standard-tier rates rising from $2 / $10 to $4 / $20 per 1M tokens, with cached input going from $0.10 to $0.20. Batch, Flex and Priority all double too. If usage stays the same, we estimate cost per index task rises from $1.62 to about $3.24.

Is Claude Opus 5.5 at max effort worth it?

For selected tasks, yes. It tops the index at 58, but at $5.98 per task it costs 3.3x as much as high effort (54, $1.82). Most teams should default to high and escalate only the hardest or highest-stakes requests.

What does Artificial Analysis's cost per task include?

AA uses provider-reported token counts where available and combines them with live measurements of each model's typical cache hit rate. The index itself is a weighted average of 10 evaluations across agents, coding, general knowledge and scientific reasoning. Treat it as an independent, comparable baseline, not a forecast of your own bill.

Sources

• Artificial Analysis: model leaderboard (Intelligence Index, cost per task, speed)
• Artificial Analysis: Intelligence Index v4.3.2 methodology
• Artificial Analysis: Kimi K3.1 Pro
• Artificial Analysis: GPT-6.1 Sol
• Artificial Analysis: Claude Opus 5.5
• Artificial Analysis: Gemini 4 Argon
• Artificial Analysis: Muse Wonder
• Artificial Analysis: DeepSeek V4.1 Flash 1017
• OpenAI API pricing
• Anthropic Claude API pricing
• Google Gemini API pricing
• DeepSeek API models and pricing

If your team needs to turn these numbers into a working model-routing setup with evals, Codersera can help you hire vetted remote AI developers who have built cost-aware LLM systems in production.