- 10
- November
"AI Models Compared, November 2026: Claude Fable 5.2, Opus 5.5, GPT-6.1 Terra, Gemini 4 Carbon, Haiku 5.5 and the Chinese Models — Which One for Which Job" — the short answer is As of 10 November 2026 the highest-scoring model on the independent Artificial Analysis index is Claude Fable 5.2 (65), followed by Gemini 4 Carbon (60) and Claude Opus 5.5 (59). The best value for most everyday work, however, is the mid-tier: Claude Sonnet 5.5, GPT-6.1 Sol and GPT-6.1 Terra score 51–55 at $1–2 per million input tokens. (For how plans and quotas work, see Claude plan limits compared.) This article updates our July 2026 comparison with current prices and scores for 20 models from 10 vendors, then answers the one question organizations actually ask: which model for which job.
In one line: AI compared, November 2026 = Fable 5.2 tops the AA index but costs the most ($8/$40) · Opus 5.5 ($4/$20) is the sweet spot for hard work · Sonnet 5.5 and GPT-6.1 Sol ($2/$10) cover everyday documents and code · GPT-6.1 Terra ($1/$5) is the cheapest model that can operate a computer · Haiku 5.5 and GPT-6 Luna ($0.10/$0.50) for high volume · Gemini 4 Argon doubles in price on 1 December · Carbon is not on general sale · Chinese models cost 3–10x less but trail by 20–30 points; Kimi K3.1 Pro's 2.8T weights went public on 4 November · still no 2026 Thai-language benchmark for this group
Every figure here was checked on 10 November 2026 against vendor pricing pages and three independent leaderboards. Scores move weekly and some vendors raise prices on 1 December, so if you are reading this later, open the references at the end and check the live numbers first. Chinese and open-weight models are named as their vendors name them.
1. The Yardstick Changed Again: Three Things to Know Before Reading the Numbers
Our September benchmark comparison used Artificial Analysis (AA) index version v4.3.2. The index is now v5.0.1. The GDPval-AA evaluation was removed, leaving nine evaluations in four equally weighted groups: knowledge work, coding, science and general. So this month's scores cannot be set directly against last month's. The clearest example is MiMo-V2.6-Pro, which scored 46 then and 30 now, for the same model.
The three yardsticks used here, and the weakness of each
- Artificial Analysis Intelligence Index v5.0.1 runs every model itself and reports by effort level. The tables use the highest effort measured, which costs more than most real use.
- Arena Text (formerly LMArena): users vote between two anonymous answers. It measures preference, not correctness, and new models are unsettled (Carbon is flagged Preliminary; Fable 5.2 was not yet listed on 9 Nov).
- ARC-AGI-2 and ARC-AGI-3 (ARC Prize Verified set): abstract reasoning on unseen puzzles. ARC-AGI-3 has a Standard column and a Provider adapter column that differ widely; this article uses Standard.
The Epoch Capabilities Index and Terminal-Bench 4.1 tables did not load on the check date, so they are omitted. Terminal-Bench 4.0 is already a component of the AA index.
2. The Big Table: 20 Models by Price, Context and Independent Score
Prices are USD per 1 million tokens, input / output, at standard rates (no batch, cache or fast mode). "Cost per task" is what AA actually paid to run its full index on that model at maximum effort. It is an independent figure that reflects how much the model writes, not just the list price. A dash means the official page does not show the figure or did not load on the check date.
| Model | Price in / out | Context | AA v5.0.1 | Cost per task (AA) | Arena 9 Nov | ARC-AGI-2 |
|---|---|---|---|---|---|---|
| Claude Fable 5.2 | $8 / $40 | 1M | 65 | $26.54 | not yet listed | 93.8% |
| Gemini 4 Carbon | not on general sale | – | 60 | $7.76 | 1526 (#1, prelim) | – |
| Claude Opus 5.5 | $4 / $20 | 1M | 59 | $10.04 | 1506 (#4) | 93.3% |
| GPT-6 Astra | $10 / $50 | 1.05M | 57 | $6.24 | 1474 (#41) | 95.0% |
| Claude Sonnet 5.5 | $2 / $10 | 1M | 55 | $10.02 | 1472 (#46) | – |
| GPT-6.1 Sol | $2 / $10 | 1.05M | 55 | $1.27 | 1482 (#27) | 94.2% |
| GPT-6.1 Terra | $1 / $5 | 1.05M | 51 | $0.96 | not yet listed | – |
| Gemini 4 Argon | $2 / $10 (→ $4 / $20 on 1 Dec) | 1M | 50 | $9.95 | 1522 (#2) | 91.3% |
| Kimi K3.1 Pro (open weights) | – | 1M | 46 | $3.68 | 1483 (#24) | – |
| Claude Haiku 5.5 | $0.10 / $0.50 (≤100K) | 1M | 36 | $0.41 | – | – |
| Grok 4.7 | – | – | 35 | $7.77 | – | – |
| Qwen3.8-Max | – | 1M | 34 | $10.02 | 1484 (#23) | – |
| DeepSeek V4.1-Flash-1017 | $0.15 / $0.60 off-peak (×2 during Thai office hours) | 1M | 33 | $0.60 | 1475 (#37) | 78.1% |
| GLM-5.3 | $1.40 / $4.40 | – | 32 | $5.01 | 1483 (#26) | – |
| MiMo-V2.6-Pro | – | – | 30 | $0.26 | 1480 (#30) | – |
| MiniMax M3.1 Flash | $0.15 / $0.60 | – | 29 | $0.49 | – | – |
| GPT-6 Luna | $0.10 / $0.50 | 1.05M | 28 | $0.15 | – | – |
| Mistral Large 4 (open weights) | – | – | 26 | $2.06 | – | – |
| Muse Spark Open (open weights) | – | – | 25 | $2.02 | – | – |
| Gemini 3.8 Flash | $0.75 / $3.75 (→ $1.50 / $7.50 on 1 Jan 2027) | – | – | – | 1497 (#9, prelim) | – |
Three things the table says plainly. First, the top group is only a few points apart (65 / 60 / 59 / 57) but 2–10x apart in price. Second, GPT-6.1 Sol and Terra score 55 and 51 at a cost per task under $1.30, about 8x cheaper than Sonnet 5.5 at the same score, because Sonnet writes far more at maximum effort. Third, the top Chinese models sit at 30–46, 20–30 points behind the leaders under the new formula, yet on Arena, which measures preference, Kimi, Qwen, GLM and MiMo rank 23–30, close to GPT-6.1 Sol.
3. The Four Claude Models Differ in Price and Speed, Not Context
All four offer a 1-million-token context and 128K output. Price differs 80x between Haiku and Fable. Fable 5.2 has a fast mode that is 2.5x quicker at 2x the price; Opus 5.5 has fast mode too. The launch details and the four breaking changes are in our Fable 5.2 article.
| Model | Standard in / out | Fast mode | Batch | AA at high / max | Use it for |
|---|---|---|---|---|---|
| Fable 5.2 | $8 / $40 | $16 / $80 | $4 / $20 | 62 / 65 | Multi-step work that must finish on its own, hard code, on-screen agents (OSWorld 89.3%, vendor figure) |
| Opus 5.5 | $4 / $20 | $8 / $40 | $2 / $10 | 52 / 59 | Hard work in general, long analysis, code, without paying double for Fable |
| Sonnet 5.5 | $2 / $10 | none | $1 / $5 | 42 / 55 | Documents, summaries, customer replies, everyday code: most of an organization's work |
| Haiku 5.5 | $0.10 / $0.50 (over 100K: $0.50 / $2.50) | none | $0.05 / $0.25 | 30 / 36 | Classification, extraction, screening, subagents, high-volume work |
If you use claude.ai rather than the API (Application Programming Interface): Fable 5.2 on Pro and Team Standard is available through usage credits only; Max, Ultra and Team Premium get it within the normal quota. That rule is unchanged from 5.1; only a first-week promotion runs to 16 November. The full table is in Claude plan limits compared, and organization pricing is at Claude pricing for organizations.
4. OpenAI and Google Got Cheaper This Year, but Google Is About to Raise Prices
OpenAI prices the GPT-6 family as a clear ladder: Astra $10/$50 for research and complex code, Sol $2/$10 for general work, Terra $1/$5 for images and computer use, Luna $0.10/$0.50 for volume. All have a 1,050,000-token context, but a request with more than 272K input tokens is charged 2x on input and 1.5x on output for the whole request. Terra's details and its system-card cautions are in our GPT-6.1 Terra article.
| Model | in / out | Context | AA (max) | ARC-AGI-2 / 3 | Verified strength |
|---|---|---|---|---|---|
| GPT-6 Astra | $10 / $50 | 1.05M | 57 | 95.0% / 62.7% | Highest ARC-AGI-2 in the table |
| GPT-6.1 Sol | $2 / $10 | 1.05M | 55 | 94.2% / 52.7% | Cost per task $1.27, best value in the 55-point group |
| GPT-6.1 Terra | $1 / $5 | 1.05M | 51 | – | computer_use + mcp in the Responses API at the lowest price in the group |
| GPT-6 Luna | $0.10 / $0.50 | 1.05M | 28 | – | Cost per task $0.15, lowest in the table |
| Gemini 4 Argon | $2 / $10 → $4 / $20 (1 Dec 2026) | 1M | 50 | 91.3% / 49.8% | Arena #2: people like its answers, but cost per task is $9.95 |
| Gemini 4 Carbon | not on general sale | – | 60 | – | Arena #1 (prelim); cyber-defence organizations only since 28 Oct |
| Gemini 3.8 Flash | $0.75 / $3.75 → $1.50 / $7.50 (1 Jan 2027) | – | – | – | Arena #9, still cheap until year end |
Dates to write down: Gemini 4 Argon doubles to $4 / $20 on 1 December 2026 and Gemini 3.8 Flash doubles on 1 January 2027, per Google's pricing page. Anyone budgeting 2027 on Gemini should use the new prices. Carbon, which scores 10 points above Argon, has no general-sale price on the pricing page.
5. Chinese and Open-Weight Models: Genuinely Cheap, but Read the Conditions
The month's big news is Moonshot publishing the Kimi K3.1 Pro weights, 2.8 trillion parameters (104B active per token), 1M context, on Hugging Face on 4 November under Moonshot's own licence rather than plain Apache or MIT. Qwen 4, which Alibaba said was in training on 22 September, had not shipped on the check date; there was only a teaser on 6 November. The policy and distillation backdrop on the Chinese side is in our China AI update, September 2026.
| Model | Official price in / out | Open weights | AA | Know before you use it |
|---|---|---|---|---|
| Kimi K3.1 Pro | Pricing page did not load prices on the check date (batch at 60%) | Yes, 4 Nov 2026 · ~1.56 TB | 46 | Needs a very large GPU server to self-host; licence has conditions for large hosted services |
| Qwen3.8-Max | Not shown on the official page (third-party aggregators only) | No (Qwen3.8-27B is open) | 34 | Supports images; Arena #23; Qwen 4 not yet released |
| GLM-5.3 / GLM-5.3-Flash | $1.40 / $4.40 · Flash $0.15 / $0.50 | – | 32 | Flash is as cheap as Haiku 5.5 but has no separate AA score |
| MiMo-V2.6-Pro | Not shown on the official page | – | 30 (was 46 under v4.3.2) | Cost per task $0.26, lowest among models scoring 30+ |
| DeepSeek V4.1-Flash-1017 | $0.15 / $0.60 off-peak · $0.30 / $1.20 peak | – | 33 | Peak hours are 01–04 and 06–10 UTC on weekdays, i.e. 08–11 and 13–17 Thai time, so Thai users pay full price during office hours |
| MiniMax M3.1 Flash | $0.15 / $0.60 | – | 29 | Same price as DeepSeek off-peak, with no time condition |
| Mistral Large 4 (v26.10) | Pricing page does not list Large 4 | Yes, Mistral Ecosystem Open Weight License | 26 | European vendor; the open-weight option for organizations that want a non-Chinese source |
| Muse Spark Open (Meta) | Depends on hosting provider | Yes, 816B / 42B active | 25 | Scores below every open Chinese model under v5.0.1 |
Before sending company data to any vendor's API, answer three questions: (1) in which country is the data processed, and how many days does the contract say it is retained; (2) will the provider sign a DPA (Data Processing Agreement) that satisfies the PDPA (Personal Data Protection Act); (3) if you self-host open weights, who maintains server security and model updates. A 10x lower price means nothing if those three go unanswered. Our guide to setting up Claude for PDPA is at Configuring Claude Team for Thai PDPA.
6. Choose by Job: The One Table Worth Pinning Above Your Desk
"First choice" is the best-value model for that job on price and independent score. "Step up when" is the condition under which paying more is justified. It is not a list of the best models, because the best model is the one that passes your own test set at the lowest price.
| Job | First choice | Step up when | Cheap / open-weight alternative |
|---|---|---|---|
| Documents, summaries, drafting, customer replies | Claude Sonnet 5.5 or GPT-6.1 Sol ($2 / $10) | Hundreds of pages to summarize across files → Opus 5.5 | Gemini 3.8 Flash until year end · GLM-5.3-Flash |
| Writing and fixing code | Opus 5.5 ($4 / $20) in Claude Code | Multi-step agent work that must finish on its own, or a large repo → Fable 5.2 (FrontierCode 49.1%, vendor figure) | Kimi K3.1 Pro self-hosted · GPT-6.1 Sol |
| Agents that use a computer / click through screens | GPT-6.1 Terra ($1 / $5) for pilots · Fable 5.2 when the job must complete itself | Work where mistakes are expensive → Fable 5.2 (ARC-AGI-3 Standard 81.2%, highest on the board) | No open-weight model performs at this level yet (DeepSeek Flash scores 0.2%) |
| Image work: reading invoices, checking products from photos | GPT-6.1 Terra or Gemini 4 Argon (until 30 Nov) | Complex reasoning needed after reading the image → Opus 5.5 / Fable 5.2 | Qwen3.8-Max supports images |
| High volume: classification, extraction, screening | Claude Haiku 5.5 or GPT-6 Luna ($0.10 / $0.50) | Accuracy below target → Sonnet 5.5 / 6.1 Sol | DeepSeek V4.1-Flash outside office hours · MiniMax M3.1 Flash · GLM-5.3-Flash |
| Must run inside the organization (on-premise) | Kimi K3.1 Pro if you have the GPUs · Mistral Large 4 for a non-Chinese source | Work the open models cannot reach → a closed vendor's API under a DPA | Muse Spark Open · Qwen3.8-27B (ARC-AGI-2 42.4%) |
| Thai as the main language | Test it yourself on 20–30 real documents before trusting anyone's number | There is still no 2026 Thai-language benchmark covering this group. The SEA-HELM board dated 18 September 2026 lists only mid-size open models (Gemma 4, Qwen 3.6, Mistral Medium 3.5, SEA-LION); no Claude, GPT-6, Gemini 4 or Kimi K3. | |
7. Price per Point of Intelligence: Where the List Price Misleads
Sonnet 5.5 and GPT-6.1 Sol carry the identical list price of $2/$10, yet what AA actually paid to run the same test suite at maximum effort differed 8x ($10.02 versus $1.27), because Sonnet spends far more tokens thinking and answering. Conversely, Fable 5.2 lists slightly below Astra on input ($8 versus $10) but costs 4x more per task ($26.54 versus $6.24) at max effort. That does not say which model is better. It says effort must be set to match the job, and that real token usage has to be read from the usage dashboard rather than estimated by multiplying list price by request count.
A simple rule that works across vendors: start any new job on a mid-tier model (Sonnet 5.5, GPT-6.1 Sol, Argon) at medium effort. Measure against your own 30–50 example test set. If it passes, stop. If not, raise effort before raising the model. Go to Fable 5.2 only for jobs Opus 5.5 at high effort still misses, and drop to Haiku 5.5 or Luna the moment a job becomes repetitive high-volume work.
8. Five Cautions Before Acting on the Tables
One: the figures in tables 2 and 4 are independent, but numbers such as Fable 5.2's FrontierCode 49.1% or OSWorld 89.3% are Anthropic's own; read them as direction. Two: AA scores in the tables are at maximum effort, which is costlier and slower than you will run; Opus 5.5 at high effort scores 52, not 59. Three: Arena measures preference. Gemini 4 Argon is #2 there despite an AA score of 50, and GPT-6 Astra is #41 despite the highest ARC-AGI-2. Neither board is wrong; they measure different things. Four: scores flagged Preliminary (Carbon with 4,282 votes against Opus 4.6's 84,424) will move a lot. Five: the AA index changed formula in September and again before this month. If an article sets scores from different months side by side, be suspicious.
9. One Vendor or Several: Suitable / Not Suitable
| Approach | Suitable for | Not suitable for |
|---|---|---|
| One vendor, 2–3 tiers by job (e.g. Haiku / Sonnet / Opus) | Organizations that need one DPA, one tax invoice and simple user management, where most work is documents and code | Jobs where that vendor clearly trails, such as cheap image work that Terra does at $1 / $5 |
| Two vendors, split by job type | Development teams with a central gateway, cost-per-task measurement and someone managing two contracts | Organizations where no one yet owns the AI data policy |
| Self-hosted open weights | Bodies whose data cannot leave the premises at all, with a team to run GPUs and server security | SMEs without an infrastructure team; jobs needing top-group scores (the best open model is at 46) |
| Chinese models via API | Experiments, internal work with no personal data, high-volume jobs outside peak hours | Customer data, government data, or any work that requires a PDPA-compliant DPA |
10. The ERP View: Models Change Monthly, the Real Data Stays Put
The tables above will need rewriting next month. What does not change is the data the models need: sales, stock, budgets and document status live in the ERP (Enterprise Resource Planning) system, which is the system of record. Whatever the vendor, the AI model is an assistant that reads and drafts from that data through a channel whose permissions can be controlled. Grand Linux connects Claude to Saeree ERP over MCP (Model Context Protocol), defining which tool reads or writes what, binding each to the user's role, and keeping an audit trail of every request. That is what lets you switch model versions without rewriting the data rules. The scope of that work is at Claude Integration & Solutions, the service overview at Claude for organizations, and the monthly context in our October 2026 AI roundup.
Conclusion
In November 2026 the top group is four models within 10 points of each other: Fable 5.2, Carbon, Opus 5.5 and Astra, of which only Opus 5.5 is on general sale at $4/$20. The $2/$10 mid-tier handles most organizational work, and Terra brings image and computer-use work down to $1/$5. Chinese models are 3–10x cheaper again and now self-hostable, but trail the leaders by 20–30 points and come with data questions to answer first. For Thai, nobody has measured this group yet; you have to measure it yourself.
The best model is not the one with the highest score. It is the cheapest one that still passes your own test set.
- Saeree ERP team perspective
References
All checked 10 November 2026. Items 1–3 are independent leaderboards; items 4–12 are vendors' official pages.
- Artificial Analysis: Intelligence Index leaderboard (v5.0.1) and index methodology
- Arena: Text leaderboard (updated 9 November 2026)
- ARC Prize: ARC-AGI-2 / ARC-AGI-3 Verified leaderboard
- Anthropic: Claude API pricing (including fast mode and batch)
- OpenAI: API pricing (GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna)
- Google: Gemini API pricing (new prices on 1 Dec 2026 and 1 Jan 2027) and Gemini models
- DeepSeek: Models & Pricing (peak hours 01–04, 06–10 UTC)
- Z.ai: GLM pricing
- Moonshot: Kimi pricing (context 1,048,576 · batch 60%) and Kimi K3.1 Pro weights on Hugging Face
- Mistral: Models overview (Mistral Large 4 v26.10, open weight)
- Alibaba Cloud Model Studio: Models (qwen3.8-max · updated 2 Nov 2026)
- AI Singapore: SEA-HELM leaderboard (18 September 2026)
Using Claude in your organization? Request a quote in Thai baht with a tax invoice
Grand Linux supplies Claude Team and Enterprise with quotes in baht, tax invoices and purchase-order support, plus integration of Claude with your back-office systems over MCP. Fill in the short form and our sales team will get back to you.
Get advice / request a quoteTel 02-347-7730 | sale@grandlinux.com
