Choosing an AI Model in 2026: Rankings, Prices, and Cost per Task

Last updated: November 5, 2026.

Gemini 4 Carbon is the intelligence leader in the November 2026 snapshot: 60 on the Artificial Analysis Intelligence Index, one point ahead of Claude Opus 5.5 at 59. Carbon is also the model most developers cannot call. Google has released it only to a subset of cyber defenders and partners in its Fairwind program. Any developer can call Opus today, and it leads the Arena Agents leaderboard at 20.14%. GPT-6.1 Sol, five points behind Carbon, costs $1.27 per task on Artificial Analysis's measurement, roughly an eighth of Opus's $10.04.

Three leaderboards now give three different answers: intelligence, agent preference, and cost per task. A selection process that reads one board can either pick a model it cannot deploy or pay several times more than necessary. This guide separates the boards, gives a request path for testing candidates, and records the failure modes that appear when a leaderboard model meets a production workload.

Prerequisites

Before comparing models, fix four inputs.

• Availability. Carbon is limited to a subset of cyber defenders first, then to additional trusted partners through the Fairwind program. Google says broader access is planned; it is not generally available now, and it is not in AIHubMix's catalog. The other six models in the first-tier table are available through it.
• A task set. Twenty real tasks with your prompts, output format, and failure criteria. Public benchmarks do not test your schema.
• A price basis. Official list prices per million tokens, at standard rates. Introductory prices expire: Carbon is $2/$10 during its launch period and $4/$20 at standard rates; Argon is $2/$10 through November 30, 2026, then $4/$20.
• A client. Python with the openai package and an AIHUBMIX_API_KEY, or the equivalent for a direct vendor endpoint. The sample below uses AIHubMix's OpenAI-compatible endpoint.

What the three boards measure

The intelligence index is a composite of reasoning, coding, knowledge, and agentic evaluations. The current board puts Carbon at 60, Opus 5.5 at 59, GPT-6 Astra at 57, and a cluster at 55 that includes Claude Sonnet 5.5 and GPT-6.1 Sol. Google's launch post cited 59 for Carbon on an earlier v5.0 snapshot; the current leaderboard shows 60. That one-point difference is a version change, not a new model.

Arena Agents measures something else: how often people prefer a model's output on agent tasks. Opus 5.5 ranks first at 20.14%, Carbon second at 16.72%, Sonnet third at 15.42%, and Astra fourth at 14.70%.

Cost per task is the third axis. Artificial Analysis runs each model through the same evaluation and records what the run cost, including reasoning and output tokens. Terra comes in at $0.96, Sol at $1.27, Astra at $6.24, Carbon at $7.76, Sonnet at $10.02, and Opus at $10.04.

A model can win one board and lose another. Carbon wins intelligence and is second on Arena, but most teams cannot use it. Opus wins Arena and is second on intelligence. Sol wins cost among models that score 55 or better.

The first tier at a glance

Model	AA II	Arena Agents	List price (in / out)	Best for
Gemini 4 Carbon	60	#2	$4 / $20	Security research
Claude Opus 5.5	59	#1	$4 / $20	Complex agents
GPT-6 Astra	57	#4	$10 / $50	Computer use
Claude Sonnet 5.5	55	#3	$2 / $10	Precise coding
GPT-6.1 Sol	55	#6	$2 / $10	High-volume agents
GPT-6.1 Terra	51	#7	$1 / $5	Value and vision
Gemini 4 Argon	50	#8	$4 / $20	Multimodal research

Prices are per million tokens. Carbon is $2/$10 during its launch period and $4/$20 at standard rates. Argon is $2/$10 through November 30, 2026, then $4/$20. Carbon is not in AIHubMix's catalog; the other six models in this table are available through AIHubMix. Artificial Analysis (AA) scores use each model's highest-scoring setting: max for Opus, Astra, Sonnet, Sol, and Terra; high for Carbon and Argon.

Two things stand out. The price spread inside the first tier is enormous: Terra's token price is a tenth of Astra's, and its cost per task is about 15% of Astra's. And the order is not stable across boards: Sonnet is third on Arena and fourth on the intelligence index; Astra is third on the index and fourth on Arena.

Opus is the production default for work that needs judgment across many steps. It scores 59 on the index and 20.14% on Arena, and Anthropic reports 66.4% on Terminal-Bench 4.0 at xhigh, 1846 Elo on GDPval-AA v2.1, and 67.7% on Humanity's Last Exam with tools. It lists at $4/$20, supports a 1M-token context window and 128K output tokens, and Artificial Analysis measures $10.04 per task at 26.2 minutes. Anthropic says typical workloads cost about 40% less than they did on Opus 5.

Haiku 5.5, outside the first tier, scores 36, lists at $0.10/$0.50 per million tokens for prompts up to 100K tokens ($0.50/$2.50 above that), supports a 1M-token context window, and costs $0.41 per task. The gap to Opus is 23 index points; use it for classification, extraction, routing, summarization, and sub-agents, not hard reasoning.

Terra is the value and vision option at $1/$5, and Artificial Analysis measures $0.96 per task. Argon is the usable Gemini 4 for multimodal work. It supports 1M tokens of context and takes text, images, audio, video, and PDFs. Google reports Vals Index first overall, 87% on ERQA, 57% on RareBench, and 52.2% on LabBench; Artificial Analysis measures $9.95 per task.

The request path

The quick start shows the current request shape. One key works across the OpenAI-compatible endpoint, a Claude-compatible endpoint (beta), and a Gemini-native endpoint, with more than 900 models on one bill. For most teams, the OpenAI-compatible endpoint is enough:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://aihubmix.com/v1",
    api_key=os.environ["AIHUBMIX_API_KEY"],
)

prompt = "Explain in three sentences why cost per task beats cost per token."

response = client.chat.completions.create(
    model="claude-opus-5-5",
    messages=[{"role": "user", "content": prompt}],
)

print(response.choices[0].message.content)

The model IDs follow the vendor names (Carbon has no ID here): claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5, claude-fable-5-1, gpt-6-astra, gpt-6.1-sol, gpt-6.1-terra, gemini-4-argon, kimi-k3.1-pro, glm-5.3, deepseek-v4.1-flash-1017, minimax-m3.1-flash, and muse-wonder. Swap the model value to test another candidate; the request shape stays the same.

If you do not want to choose per request, the auto router supports four strategies: cost_optimized (the default), balanced, quality_first, and latency_critical. Use it for exploratory work; pin critical tasks to a fixed model and keep a fallback.

Token price is not task cost

Model	AA II	AA cost per task	Bill at list prices
Gemini 4 Carbon	60	$7.76	$80
Claude Opus 5.5	59	$10.04	$80
GPT-6 Astra	57	$6.24	$200
Claude Sonnet 5.5	55	$10.02	$40
GPT-6.1 Sol	55	$1.27	$40
GPT-6.1 Terra	51	$0.96	$20
Gemini 4 Argon	50	$9.95	$80
Claude Haiku 5.5	36	$0.41	$2

The bill column is a worked example, not an AA figure: 10M input tokens plus 2M output tokens at standard list prices, assuming each request stays under its provider's long-context threshold, with no caching and no batch discount. Carbon and Argon are shown at standard rates; their launch and intro prices cut those bills to $40 each.

Three conclusions follow.

First, the same token price can produce roughly an 8x difference in cost per task. Sonnet and Sol both list at $2/$10. Sonnet's measured cost per task is $10.02; Sol's is $1.27. The difference is output length and thinking time, not the rate card.

Second, an expensive model is not always the expensive choice. Astra's token price is 2.5x Opus's, but its per-task cost is $6.24 versus Opus's $10.04, because it spends fewer tokens per task. A high rate card and a high bill are different things.

Third, context is not free. OpenAI's long-context tier applies higher rates to the full request once it passes 272K tokens: $20 input and $75 output for Astra, $4 and $15 for Sol. A 1M-token window is useful; filling it on every request is a budget decision.

Failure modes and fixes

The top score belongs to a model you probably cannot call

Carbon scores 60 but is limited to a subset of cyber defenders and Fairwind partners. A benchmark run against it is not a plan. Check access before you spend time on evaluation. For most teams, Argon is the Gemini 4 model to test.

The same token price can produce a different bill

Sonnet 5.5 and Sol both list at $2/$10. Artificial Analysis measures $10.02 per task for Sonnet and $1.27 for Sol, a 7.9x gap, and 30.6 minutes versus 20.9 minutes per task. The gap comes from how much each model writes and thinks. Measure tokens per task on your workload, not the rate card.

A higher token price can be the cheaper task

Astra lists at $10/$50, 2.5x Opus's $4/$20, but Artificial Analysis measures $6.24 per task against Opus's $10.04. In OpenAI's software-maintenance write-up, the Ultra effort configurations tell the same story: Astra spent $109.06 to produce 142 findings ($0.77 per finding), and Sol spent $15.12 to produce 131 findings ($0.12 per finding): 92% of the findings at 14% of the API cost. A higher rate card does not imply a higher bill, so measure tokens per task before assuming the higher-priced model is the expensive choice.

Long context changes the rate, not just the window

Above 272K input tokens, OpenAI applies higher rates to the full request: $20 input and $75 output for Astra, $4 and $15 for Sol. A 1M-token context window is capacity, not a free pass. Keep requests under the threshold unless the task needs the whole document.

The benchmark comparison is not effort-matched

Sonnet's vendor-reported Terminal-Bench 4.0 score of 70.6% is higher than Opus's 66.4%, but Sonnet's effort setting is not published, while Opus's figure is at xhigh. A benchmark measures a task, not judgment. Match the effort setting and the harness before treating a score as a ranking. GLM's Terminal-Bench 4.0 score of 41.9% is its Max configuration on the mini-swe-agent harness.

Open weights are close on coding, not on general intelligence

Kimi K3.1 Pro reports 65.7% on Terminal-Bench 4.0, 0.7 points behind Opus 5.5, and supports a 1,048,576-token context window. Artificial Analysis puts Kimi at 46 on the intelligence index, 13 points below Opus, and outside the top ten on Arena. A model can be close on one coding benchmark and far behind on the mix of reasoning, knowledge, and agentic tasks that the index measures. Use Kimi for terminal coding, CAD and 3D work, and financial documents where an open license or self-hosting matters.

Open weights: what the license buys

Model	Size and license	Context	List price (in / out)	Why it matters
Kimi K3.1 Pro	2.8T / 104B; Kimi license	1M	Not published	Terminal-Bench 4.0 65.7%
GLM 5.3	753B MoE; glm-5.3 license	1M	$1.40 / $4.40	Terminal-Bench 4.0 41.9% (Max)
DeepSeek V4.1 Flash 1017	763B; MIT	1M	from $0.15 / $0.60	Vision plus 1M context
MiniMax M3.1 Flash	428B / 23B active; community	1M	¥1 / ¥4	OSWorld 2.0 61.7%

Benchmark figures in this table are vendor-reported; GLM's Terminal-Bench 4.0 score is its Max configuration on the mini-swe-agent harness. Moonshot's pricing page did not show a USD list rate in the version checked; Artificial Analysis measures $3.68 per task. GLM's official price is $1.40/$4.40; AIHubMix lists it below list at $1.13/$3.94 per million tokens. DeepSeek's official price is $0.15/$0.60 off-peak and $0.30/$1.20 at peak; peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Cache-hit input costs $0.003 off-peak and $0.006 at peak per million tokens. MiniMax prices in CNY at ¥1/¥4 per million tokens.

Kimi is the headline: a 65.7% Terminal-Bench score within 0.7 points of Opus is a real achievement, and the 1M context makes it usable for repository-scale work. The intelligence gap is the other half: 46 versus 59. GLM 5.3 is the cost-oriented choice; AA puts it at 32 at max effort, well below Kimi's 46, but it is a 753B MoE under the glm-5.3 license with a vendor-reported 41.9% on Terminal-Bench 4.0 (Max, mini-swe-agent), 1769 Elo on GDPval-AA v2, and 62.5% on Humanity's Last Exam (HLE) with tools. DeepSeek V4.1 Flash 1017 is the cost and ownership play: 763B under MIT, 1M context, 384K max output, and vision input. MiniMax M3.1 Flash is the multimodal agent option: 428B total parameters, 23B active per token, 1M context, text/image/video input, and an OSWorld 2.0 score of 61.7.

The value of open weights is not that the gap has closed everywhere. It is that you can own the model when data cannot leave your network, when costs must be fixed, or when you need to fine-tune. On those constraints, a few index points matter less than deployment freedom.

How to choose

If you need...	Start with	Then consider
One flagship for everything	Claude Opus 5.5	Carbon if you have access
Daily coding and documents	Claude Sonnet 5.5	Sol for volume
Large-scale, cost-sensitive agents	GPT-6.1 Sol	Terra or DeepSeek
Multimodal and long video	Gemini 4 Argon	Muse Wonder
Open weights or self-hosting	Kimi K3.1 Pro for capability	GLM 5.3 or DeepSeek for cost
High-frequency small tasks	Claude Haiku 5.5	Terra

Muse Wonder is the other multimodal agent in the Arena top ten, ranked ninth at 9.42%; its Full configuration scores 50 on AA, while the standard configuration scores 44.

Three rules make the table work.

Run 20 real tasks before committing. Use your prompts, your output format, and your failure criteria, not a public benchmark. Twenty tasks are enough to expose a model that cannot follow your schema or that folds under a long context.

Split decisions from execution. Let Opus or Sonnet plan, review, and handle the hard cases; let Haiku, Sol, or DeepSeek execute the well-specified steps. In an agent system, the most expensive model does not need to do every step.

Keep a replaceable layer. Models change every few months. A single API, a single bill, and a single log matter more than picking the winner of one month.

What cannot be reproduced or verified

• The intelligence index and per-task figures are Artificial Analysis measurements. They depend on AA's evaluation setup and token accounting; this guide does not reproduce them.
• Vendor benchmark scores are self-reported. Opus's 66.4% Terminal-Bench 4.0 is at xhigh; Sonnet's 70.6% has no published effort setting. Carbon's 88.2% SEC-bench Pro, 71.0% CWE-bench, and 54.8% ExploitGym come from Google's launch post. Kimi's 65.7% comes from Moonshot. Harnesses differ: GLM's 41.9% is its Max configuration on mini-swe-agent.
• Carbon's access is restricted, and Google's launch post does not list a context window for it. The post cites 59 on AA v5.0 while the current leaderboard shows 60, a version change.
• Kimi's USD list price was not published on Moonshot's pricing page in the version checked.
• The bill column is a worked example, not an AA figure: 10M input tokens plus 2M output tokens at standard list prices, with each request under its provider's long-context threshold, no caching, and no batch discount. Carbon and Argon are shown at standard rates; their launch and intro prices cut those bills to $40 each.
• Prices and access change. This guide is dated November 5, 2026.

Sources

1. Artificial Analysis Intelligence Index leaderboard, accessed November 5, 2026.
2. Arena Agents leaderboard, accessed November 5, 2026.
3. Frontier vendor material: Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 4 Carbon, Gemini 4 Argon, Gemini API models, OpenAI model catalog, GPT-6 family maintenance post, Kimi K3.1 Pro, Kimi pricing.
4. Open-weight vendor material: GLM 5.3 pricing, GLM 5.3 model card, DeepSeek pricing, DeepSeek V4.1 Flash 1017 model card, MiniMax pricing, MiniMax M3.1 Flash model card.
5. AIHubMix: model catalog and API quick start, accessed November 5, 2026.