Back to the trend of the day: Intel's boss warns billions of AI helpers will be hard to control

Poly Rating · from the trend of the day

AI Agent Cost Comparison 2026: Most Work per Dollar

On Arena's Nov 4 coding board, a median Claude Opus 5.5 task costs $17.45 and GPT-6.1 Terra $1.40: about 12 times the price for 23.38% net improvement against 13.76%.

No. 1: GPT-6.1 Sol (OpenAI Codex)
10 places
Current as of November 4, 2026, 19:18 EST

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

1
GPT-6.1 Sol (OpenAI Codex)
The cheap pick three separate boards agree on: 12.51% Arena net improvement at $0.98, 62.9% on the Coding Agent Index at $1.04, 58.2% on Terminal-Bench at $1.92.
Included it: ChatGPT, Meta, Grok, StepFun, Kimi, DeepSeek. Did not include: Nemotron, Wenxin.
In their lists: ChatGPT: 2 · Meta: 1 · Grok: 3 · StepFun: 3 · Kimi: 1 · DeepSeek: 4
What each AI noted 6
- ChatGPTplace 2
  A $0.98 median task cost and 10.3-minute median duration make it a practical value candidate for controlled coding workflows.
  Source ↗
- Metaplace 1
  $0.98 median in Arena, $1.27 per Intelligence Index task for 55 pts – 8x cheaper than Opus 5.5 max at $10.04 for 58 pts
- Grokplace 3
  A 98-cent overall median buys 12.51% net improvement, a step ahead of Terra and a step behind on shell recovery.
  Source ↗
- StepFunplace 3
  OpenAI's cost play: within two points of its pricier siblings on the same board at $0.56 a median task.
- Kimiplace 1
  62.9% of real engineering tasks solved at $1.04 and 15.5 minutes each — about 60 index points per dollar, six times the work-per-dollar of the leaderboard's top scorer.
  Source ↗
- DeepSeekplace 4
  58.2% for $1.92 — one-fifth of Astra's $9.90 on the same test.
6 AIs put it here
Source ↗
2
Claude Opus 5.5 (Claude Code)
Highest overall net improvement (20.14%), but $17.45 and 39 minutes per median coding task. Terminal-Bench runs price it at $14.30–$19.07.
Included it: ChatGPT, Meta, Grok, StepFun, Kimi, DeepSeek. Did not include: Nemotron, Wenxin.
In their lists: ChatGPT: 6 · Meta: 8 · Grok: 7 · StepFun: 4 · Kimi: 9 · DeepSeek: 6
What each AI noted 6
- ChatGPTplace 6
  It leads Arena's overall November 3 ranking, but $6.80 per median task means its premium must be justified by fewer failed or repeated attempts.
  Source ↗
- Metaplace 8
  $0.08 median absolute cheapest – 0.62% overall, 1.9 min, 26.1K tokens at $0.10/$0.50
  Source ↗
- Grokplace 7
  Most net improvement on the board, 20.14%, at $6.79 overall and $17.45 when the task is code.
  Source ↗
- StepFunplace 4
  The best value inside the best family: #1 steerability and #2 confirmed success at $1.31 a task, 64% below its own Max tier.
  Source ↗
- Kimiplace 9
  The priciest reliable workhorse: 66% task success at $13.04 — you pay roughly double per task versus Codex, and independent tracking says it finishes in a 2.2-minute median.
  Source ↗
- DeepSeekplace 6
  64.8% at $14.30 — reliable work, at a price that needs a high-value task to pay back.
6 AIs put it here
Source ↗
3
Claude Sonnet 5.5 (Claude Code)
Ties Opus 5.5 on confirmed coding success (24.98%) at $8.39 per median coding task instead of $17.45. Overall: 15.42% net improvement at $3.86.
Included it: ChatGPT, Meta, Grok, StepFun, Kimi. Did not include: Nemotron, DeepSeek, Wenxin.
In their lists: ChatGPT: 5 · Meta: 5 · Grok: 5 · StepFun: 6 · Kimi: see list
What each AI noted 5
- ChatGPTplace 5
  The $3.86 median cost is much higher than the sub-dollar alternatives, but 19.95% confirmed success makes its reliability-adjusted value competitive.
  Source ↗
- Metaplace 5
  74.55% GAIA leader for $178.20 – 15.42% Arena confirmed for $3.83, 16.4 min – best research work per dollar
- Grokplace 5
  Ties Opus on confirmed coding success, 24.98%, at $8.39 instead of $17.45.
  Source ↗
- StepFunplace 6
  Second-best confirmed success on the whole board (20.21%) for $3.86 a task — the mid-tier Anthropic pick that keeps most of the reliability.
  Source ↗
- Kimi
  Named differently in its own list: open the full list.
  see list
5 AIs put it here
Source ↗
4
Gemini 4 Argon
On chat tasks, 13.22% net improvement for 52 cents in 3.8 minutes. Weaker on code: 8.54% at $2.14 on Arena, and $5.84 per task in Google's Antigravity CLI.
Included it: ChatGPT, Meta, Grok, Kimi. Did not include: Nemotron, StepFun, DeepSeek, Wenxin.
In their lists: ChatGPT: 3 · Meta: 3 · Grok: 2 · Kimi: 3
What each AI noted 4
- ChatGPTplace 3
  At $0.97 and 7.3 minutes per median task, it is faster and cheaper than Carbon, though Carbon has the stronger overall improvement signal.
  Source ↗
- Metaplace 3
  $0.97 median, 10.53% overall – 0.42% tool hallucination lowest among top 10, 13.52% steerability
- Grokplace 2
  On chat tasks, 13.22% net improvement for 52 cents and 3.8 minutes.
  Source ↗
- Kimiplace 3
  Best DeepSWE score on the board (78.8%) at $5.84 a task — the strongest hard-coding work per dollar among the big-three harnesses.
  Source ↗
4 AIs put it here
Source ↗
5
Gemini 4 Carbon
Most quality for under $2: 16.72% net improvement and 18.02% confirmed success at a $1.69 median, about a quarter of Opus 5.5's $6.79. Arena still marks it preliminary.
Included it: ChatGPT, Meta, Grok, StepFun. Did not include: Nemotron, Kimi, DeepSeek, Wenxin.
In their lists: ChatGPT: 1 · Meta: 4 · Grok: 4 · StepFun: 2
What each AI noted 4
- ChatGPTplace 1
  At $1.69 per median task, it delivers 18.04% user-confirmed success in Arena's November 3 snapshot, making it the strongest observed balance of price and completion among leading models.
  Source ↗
- Metaplace 4
  16.72% confirmed #2 overall for $1.69 – 19.32% steerability best, 0.51% hallucination
- Grokplace 4
  The value leader that still wins tasks: 16.72% overall net improvement at $1.69, and Arena marks it preliminary.
  Source ↗
- StepFunplace 2
  The efficiency pick that also wins: #2 overall on Agent Arena at $1.69–$1.80 a task, a fifth of the price of the agent above it on confirmed success.
  Source ↗
4 AIs put it here
Source ↗
6
DeepSeek V4.1 Flash
The cheapest real entry at a $0.16 median, but only 2.37% net improvement on Nov 4; an earlier Arena post showed +4.87% at $0.07.
Included it: ChatGPT, Meta, Grok, StepFun. Did not include: Nemotron, Kimi, DeepSeek, Wenxin.
In their lists: ChatGPT: 8 · Meta: 7 · Grok: 9 · StepFun: 1
What each AI noted 4
- ChatGPTplace 8
  At $0.16 per median task, it is among the cheapest observed options, but the 3.38% confirmed-success signal prevents a higher finished-work ranking.
  Source ↗
- Metaplace 7
  $0.16 median cheapest coding – 2.37% overall, 86.5K tokens, $0.15/$0.60 per M
  Source ↗
- Grokplace 9
  Close to Astra on quality, 13.86% versus 14.70%, at a similar $6.40 median and a $16.30 coding median.
  Source ↗
- StepFunplace 1
  $0.07 median per task for +4.87% net improvement — roughly 97 times cheaper than the board's #1 agent for work that is only a few points worse.
  Source ↗
4 AIs put it here
Source ↗
7
Devin (Cognition)
The cost per task depends on who counts it: $0.60 (vendor claim), $1.40 at 70.4% (third-party DeepSWE), $7.90 (Coding Agent Index), about $9–10 per merged task (third-party estimate).
Included it: Meta, StepFun, Kimi, DeepSeek. Did not include: ChatGPT, Grok, Nemotron, Wenxin.
In their lists: Meta: see list · StepFun: 7 · Kimi: 5 · DeepSeek: 2
What each AI noted 4
- Meta
  Named differently in its own list: open the full list.
  see list
- StepFunplace 7
  Cognition claims 64% lower cost than the frontier coding model — implying about $1.18 a task — while other arithmetic puts a merged task near $9–$10.
  Source ↗
- Kimiplace 5
  $7.90 buys 61.7% task completion with SWE-2 Medium handling the heavy lifting — mid-pack price, top-tier reliability per dollar for delegated tickets.
  Source ↗
- DeepSeekplace 2
  68.8 at $0.60 — if you count the FrontierCode run and not the $500/month seat.
4 AIs put it here
Source ↗
8
GPT-6.1 Terra
The cheapest strong performer: a $0.79 median for 11.40% net improvement, and $1.40 per coding task against $17.45 for Opus 5.5.
Included it: ChatGPT, Meta, Grok. Did not include: Nemotron, StepFun, Kimi, DeepSeek, Wenxin.
In their lists: ChatGPT: 4 · Meta: 2 · Grok: 1
What each AI noted 3
- ChatGPTplace 4
  The $0.79 median task cost undercuts Sol's $0.98, but the observed overall improvement signal is lower.
  Source ↗
- Metaplace 2
  Cheapest top-10 scorer at $0.96 per task for 48 pts – $0.79 median Arena, 11.4% success, 15.18% steerability
- Grokplace 1
  11.40% overall net improvement at a 79-cent median, and $1.40 on a coding task.
  Source ↗
3 AIs put it here
Source ↗
9
Kimi K3.1 Pro
A $1.47 median for 5.29% net improvement (preliminary): under half of Terra's gain at nearly twice the price. Toloka found 41% of K3's tasks solved only sometimes.
Included it: Meta, Grok, StepFun. Did not include: ChatGPT, Nemotron, Kimi, DeepSeek, Wenxin.
In their lists: Meta: 6 · Grok: 6 · StepFun: 5
What each AI noted 3
- Metaplace 6
  #1 open model tie – 5.29% overall, $1.47 median, 6.8 min, 41.4K tokens at $2/$10 per M
  Source ↗
- Grokplace 6
  Positive, not cheap-and-empty: 5.29% overall net improvement at $1.47, far behind Terra's 11.40% at 79 cents.
  Source ↗
- StepFunplace 5
  $1.03 a median task and it still lands just off the Pareto frontier — the cheapest credible route into top-tier agentic work.
3 AIs put it here
Source ↗
10
GPT-6 Astra
Fourth on overall quality (14.70%), yet $6.16 per median task and $15.30 on code; Sonnet 5.5 gets more done for $3.86.
Included it: ChatGPT, Grok. Did not include: Meta, Nemotron, StepFun, Kimi, DeepSeek, Wenxin.
In their lists: ChatGPT: 10 · Grok: 8
What each AI noted 2
- ChatGPTplace 10
  Its $6.16 median task cost and 14.28% confirmed-success signal make it a premium option rather than a general price-performance winner.
  Source ↗
- Grokplace 8
  Fourth on overall quality, 14.70%, and already $15.30 before a coding task is done.
  Source ↗
2 AIs put it here
Source ↗

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The numbers come from different tests: Arena overall and coding medians, Terminal-Bench, and the Coding Agent Index. Costs per task differ between them, so compare them only within one test.

1GPT-6.1 Sol (OpenAI Codex)

No. 1 — GPT-6.1 Sol (OpenAI Codex). On the lists of 6 of 8 AIs; median place there: 2.5. Three separate boards agree it gets solid results for the least money, about a dollar per task.

Did you know

GPT-6.1 Sol can read a huge amount at once. Its listed context window is 1,050,000 tokens, and it can write up to 128,000 tokens in one reply. That is enough room to hold a very large codebase in a single session.
Source ↗

What it is: OpenAI's Codex coding agent running the GPT-6.1 Sol model.
Its edge: Costs far less per task than the two Claude options on every board.
Who it suits: Suits budget-minded coding work; less suited if you need the highest net improvement.

Net improvement versus cost12.51% net at $0.98

Wikipedia · Wikidata

Why it is here

On Terminal-Bench Science 0.1 at maximum effort, GPT-6.1 Sol averages $5.47 per task. Claude Opus 5.5 averages $23.21 on the same harness, so Sol costs about a quarter as much.
Source ↗
TechRepublic's October 2026 comparison reports a cost per task of $1.04 for Codex with GPT-6.1 Sol. Claude Code with Sonnet 5.5 is listed at $14.19. I could not open this page to check the figures myself; they come from the search summary.
Source ↗

Did you know

Its listed API price is $2 per million input tokens and $10 per million output tokens. That is the price that sits behind the 'cheap pick' label, since most of an agent's cost comes from the tokens it reads and writes.
Source ↗

History

In May 2025 the cloud version of Codex launched, but few people outside OpenAI kept using it. The team, which had about seven weeks to build it, then turned to tools that run on your own computer. That change became the base for today's Codex.
Source ↗

Where to go further

Open OpenAI's official GPT-6.1 Sol launch post. Read the benchmark table first, then check the plan notes to see which Codex plans include the model before you try it.
Source ↗

Video: OpenAI Codex Explained for Normal People (“Claude Cowork Killer”) · Paul J Lipsky

Compare the top three ↓

2Claude Opus 5.5 (Claude Code)

No. 2 — Claude Opus 5.5 (Claude Code). On the lists of 6 of 8 AIs; median place there: 6.5. It gets the most net improvement overall, but each coding task costs far more and takes much longer.

WOTW2026 · CC0 · Wikimedia Commons

Did you know

Claude Code's rewind does not undo everything. It only tracks edits made with Claude's own file-editing tools. Changes made by shell commands, such as deleting or moving files, are not covered. Checkpoints do survive across sessions.
Source ↗

What it is: Anthropic's Claude Code agent running the Opus 5.5 model.
Its edge: Highest net improvement on the board, though it is the most expensive per task.
Who it suits: Suits hard, high-value work; less suited to easy tasks or tight budgets.

Net improvement versus cost20.14% net, $17.45 per task

Wikipedia · Wikidata

Why it is here

Artificial Analysis lists Claude Code with Opus 5.5 at 62 on its Coding Agent Index. That is the cost-heavy side of the trade: $11.42 per task and 1.0 hour per task.
Source ↗
On Terminal-Bench Science 0.1, Claude Opus 5.5 averages $23.21 per task. GPT-6.1 Sol averages $5.47 on the same test, so Opus costs over four times as much.
Source ↗

History

Claude Code started in February 2025 as a limited research preview, released alongside the Claude 3.7 Sonnet model.
Source ↗

Keep in mind

With unclear intent or missing context, auto mode may still approve unsafe actions, and may block harmless ones. Each action gets an extra review step, which can add delay and cost more tokens over time.
Source ↗
Anthropic recommends running auto mode in isolated environments, especially for sensitive code. In practice, treat it as a helper to watch, not something to leave alone.
Source ↗

Where to go further

Open the permission modes page to see what each mode lets Claude Code do without asking. Read the Auto mode part first, since it is now the starting mode.
Source ↗

Video: Every Claude Code Concept Explained for Normal People · Simon Scrapes

Compare the top three ↓

3Claude Sonnet 5.5 (Claude Code)

No. 3 — Claude Sonnet 5.5 (Claude Code). On the lists of 5 of 8 AIs; median place there: 5. It matches Opus on confirmed coding success at about half the price per task, so it keeps most of the value.

WOTW2026 · CC0 · Wikimedia Commons

What it is: Anthropic's Claude Code agent running the cheaper Sonnet 5.5 model.
Its edge: Ties Opus on confirmed coding success for much less money per task.
Who it suits: Suits Claude users who want reliability on a budget; less suited to the very cheapest work.

Net improvement versus cost15.42% net at $3.86

Wikipedia · Wikidata

Why it is here

Artificial Analysis lists Claude Sonnet 5.5 as a Claude Code variant next to Opus 5.5. The Opus 5.5 setup is shown at $11.42 per task and 1.0 hour per task. This page does not give Sonnet's own cost figure.
Source ↗
On AutomationBench, Claude Sonnet 5.5 averaged $1.14 per task, according to Vellum's write-up. That is a different test from coding, so it only shows that Sonnet can be cheap on some jobs.
Source ↗

Video: Every Claude Code Concept Explained for Normal People · Simon Scrapes

Compare the top three ↓

The top three side by side

	GPT-6.1 Sol (OpenAI Codex)	Claude Opus 5.5 (Claude Code)	Claude Sonnet 5.5 (Claude Code)
In the AIs' lists	6 / 8	6 / 8	5 / 8
Net improvement versus cost	12.51% net at $0.98	20.14% net, $17.45 per task	15.42% net at $3.86
Put first by	Meta, Kimi	—	—

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

03Worth knowing

What the AIs noticed while building the list.

A coding task and a chat task are different purchases: Opus 5.5's median is $17.45 and 39.3 min on code against $3.68 and 10.4 min on chat; Terra's is $1.40 and $0.58.

Noticed by:

Included it: Grok. Did not include: ChatGPT, Meta, Nemotron, StepFun, Kimi, DeepSeek, Wenxin.

01Final list

GPT-6.1 Sol (OpenAI Codex)

Claude Opus 5.5 (Claude Code)

Claude Sonnet 5.5 (Claude Code)

Gemini 4 Argon

Gemini 4 Carbon

DeepSeek V4.1 Flash

Devin (Cognition)

GPT-6.1 Terra

Kimi K3.1 Pro

GPT-6 Astra

02More about the top three

1GPT-6.1 Sol (OpenAI Codex)

Did you know

2Claude Opus 5.5 (Claude Code)

Did you know

3Claude Sonnet 5.5 (Claude Code)

The top three side by side

More ratings

03Worth knowing

04How we counted

1The question

2Why this topic

3Where it applies

4What we looked at

5Who ranked

6How the places were set

05Each AI's own list

What ChatGPT noticed

What Meta noticed

What Grok noticed

What StepFun noticed

What Kimi noticed

What DeepSeek noticed