Best in task.

One row per job, from public tests. Open a row to see every result, its price, and what to check before you use it.

Best resultMost Efficient OptionFree to run in-house
Read Supplier DocumentsPull the fields you need from a technical data sheet, a safety data sheet, or a certificate of analysis.OXR-1.0 Specialist · 1.3B97.03No prices publishedTeleOCRSpecialist · 1.2BFree to download96.91

Every order, spec check, and compliance file starts with a supplier document. Reading it wrong means a wrong price, a wrong grade, or a missed hazard.

The six best readers are small models you can download and run on your own computers. The top one needs a separate license at most chemical companies.

OXR-1.0 joined the test on 29 October and scores 0.12 above TeleOCR. The test publishes no margin of error, so the two are level and OXR-1.0 is named first. OXR-1.0's license allows free business use only while a company has no more than $2 million in revenue and no more than $2 million in funding; above that, you need a separate license from SoMark. TeleOCR is the best result most chemical companies can run on their own servers without one. An outside tester challenged TeleOCR's table scores on 26 October, and OmniDocBench has not ruled on it yet.

What to check

Run it on 20 of your own supplier documents, including a scan and a multi-page safety data sheet, and check every extracted value against the page.

Read the highlightTeleOCR: A Free Model That Tops One Document TestRead the highlightClaude Haiku 5.5: Strong for Its Price, Weak on Documents

Whole-document reading score: text, tables, and formulas, out of 100.

  1. OXR-1.0Specialist · 1.3B97.03
  2. TeleOCRSpecialist · 1.2B · Free to download96.91
  3. OvisOCR2Specialist · 0.8B · Free to download96.47
  4. PaddleOCR-VL-1.6Specialist · 0.9B · Free to download96.34
  5. MinerU2.5-ProSpecialist · 1.2B · Free to download · The test's own product95.75
  6. GLM-OCRSpecialist · 0.9B · Free to download95.22
  7. Gemini 3 ProGoogle92.91
  8. Gemini 3 FlashGoogle92.62
  9. GPT-5.2OpenAI86.59

OmniDocBench ↗Run by OpenDataLab (Shanghai AI Lab), which also makes MinerU, a model it ranks. Board updated 12 Nov 2026 · Vailent checked 13 Nov 2026

A second test puts its own maker's tool first. That tool is marked.

  1. LlamaParse Agentic PlusTool · The test's own product86.695.63 US cents a page
  2. LlamaParse AgenticTool · The test's own product82.831.25 US cents a page
  3. Pulse Ultra 2Tool81.61.50 US cents a page
  4. Upstage Document Parse 2 (Enhanced)Tool80.643.00 US cents a page
  5. Claude Opus 5.5Anthropic78.86.13 US cents a page
  6. rakedoc-nanoSpecialist · 1.16B · Free to download77.23No price published
  7. Gemini 3 FlashGoogle73.982.41 US cents a page
  8. GPT-6.1 TerraOpenAI73.641.09 US cents a page
  9. Claude Sonnet 5.5Anthropic69.212.74 US cents a page
  10. GPT-6 SolOpenAI68.853.47 US cents a page

ParseBench ↗Run by LlamaIndex, which sells LlamaParse, a product it ranks. Board updated 14 Nov 2026 · Vailent checked 15 Nov 2026

Pull Tables and Price ListsRead a supplier's multi-page price list into rows you can use.TeleOCR Specialist · 1.2BFree to download96.82No prices publishedSame as the best resultFree to download and run in-house

Price lists and spec tables carry the numbers you quote from. One shifted column is a wrong price on a customer quote.

A free 1.2B model is at the top on tables on the independent test, level with OXR-1.0. On the priced test, a Gemini setting comes within 5 points of the leader at a ninth of the price.

OXR-1.0 joined the test on 29 October and scores 0.21 below TeleOCR on tables. The test publishes no margin of error, so the two are level and TeleOCR is named first. OXR-1.0 needs a separate license from SoMark at most chemical companies. An outside tester challenged TeleOCR's table scores on 26 October, and OmniDocBench has not ruled on it yet.

What to check

Test your longest price lists. Check merged cells, footnotes, and where rows land across a page break.

Read the highlightClaude Sonnet 5.5: At the Top on Regulations, in a CrowdRead the highlightClaude Opus 5.5: First Place, and What It CostsRead the highlightTeleOCR: A Free Model That Tops One Document Test

Table structure score, out of 100.

  1. TeleOCRSpecialist · 1.2B · Free to download96.82
  2. OXR-1.0Specialist · 1.3B96.61
  3. PaddleOCR-VL-1.6Specialist · 0.9B · Free to download94.76
  4. OvisOCR2Specialist · 0.8B · Free to download94.58
  5. MinerU2.5-ProSpecialist · 1.2B · Free to download · The test's own product93.42
  6. GLM-OCRSpecialist · 0.9B · Free to download92.83
  7. Gemini 3 FlashGoogle89.29
  8. Gemini 3 ProGoogle89.15

OmniDocBench ↗Run by OpenDataLab (Shanghai AI Lab), which also makes MinerU, a model it ranks. Board updated 12 Nov 2026 · Vailent checked 13 Nov 2026

This second test shows prices. It scores the small models above much lower: TeleOCR reaches 72.77 on tables. Its maker sells one of the tools, which is marked.

  1. Claude Opus 5.5Anthropic94.266.13 US cents a page
  2. LlamaParse Agentic PlusTool · The test's own product93.275.63 US cents a page
  3. GPT-6 AstraOpenAI93.189.87 US cents a page
  4. GPT-6.1 TerraOpenAI93.071.09 US cents a page
  5. GPT-6 SolOpenAI92.653.47 US cents a page
  6. oi-parserTool92.62No price published
  7. Claude Fable 5.1Anthropic91.5316.05 US cents a page
  8. Gemini 3 FlashGoogle91.192.41 US cents a page
  9. Claude Sonnet 5.5Anthropic91.072.74 US cents a page
  10. Gemini 3 Flash (minimal thinking)Google89.640.65 US cents a page

ParseBench ↗Run by LlamaIndex, which sells LlamaParse, a product it ranks. Board updated 14 Nov 2026 · Vailent checked 15 Nov 2026

Read Scans and PhotosRead a photographed delivery note or a scanned certificate.OvisOCR3 Specialist · 0.85BFree to download87.4No prices publishedSame as the best resultFree to download and run in-house

Paper still arrives: delivery notes, stamped certificates, photos from the warehouse floor. A reader that fails on a photo sends a person back to typing.

OvisOCR3, a free model you can run on your own computers, leads on photographed pages, 2.3 points above Gemini 3 Pro Preview. The test publishes no margin of error. The newest Claude, GPT, and Gemini models have not been tested here.

What to check

Photograph ten real documents on the warehouse floor, in bad light, and check every value.

Reading score on photographed pages, out of 100.

  1. OvisOCR3Specialist · 0.85B · Free to download87.4
  2. Gemini 3 Pro PreviewGoogle85.1
  3. MonkeyOCRv2-BSpecialist · 0.9B · Free to download · The test's own product81.7
  4. Kimi K3Moonshot AI · Free to download81.2
  5. dots.mocrSpecialist · 3B · Free to download77.2
  6. Chandra OCR 2Specialist · 5.3B77.1
  7. PaddleOCR-VL-1.6Specialist · 0.9B · Free to download76.3
  8. Claude Sonnet 4.6Anthropic69.3
  9. GPT-5.2OpenAI63

MDPBench ↗Run by the OCRBench authors' lab, which also makes MonkeyOCR, a model it ranks. Board updated 6 Nov 2026 · Vailent checked 9 Nov 2026

Find a Figure in Your Scanned RecordsFind what you invoiced a customer for a grade in 2014, in a scanned PDF somewhere on the shared drive.Claude Opus 5.5 Anthropic64.8%DeepSeek V4.1 FlashDeepSeekFree to download37.1%6x cheaperDeepSeek V4.1 FlashDeepSeekFree to download37.1%

Years of invoices, contracts, and certificates sit on shared drives as scans nobody can search. An assistant that finds the right file and reads one number from it saves an afternoon, but only if the number is right.

The best agent found the right number about two times in three. Most misses were a wrong number, not a blank, and the most common cause was one digit misread from the scan.

Each agent tried each of 70 questions three times. The study finds Claude Opus 5.5 not significantly ahead of Gemini 4 Argon, and by the guide's margin rule Opus is also level with GPT-6.1 Sol and Gemini 3.8 Flash. Sol costs $0.62 an attempt against Opus's $1.56. DeepSeek V4.1 Flash, an open model, gives the most score for each dollar at $0.25 an attempt and is the best result you can run on your own servers. The costs leave out the document reading service.

What to check

Make it name the file and the page for every figure, and check the figure against the page before you use it. The test is US Treasury records, not your invoices.

Attempts that found the right figure, within a small rounding range, searching 1,435 scanned US Treasury records.

  1. Claude Opus 5.5Anthropic64.8%$1.56 an attempt
  2. Gemini 4 ArgonGoogle60%$2.45 an attempt
  3. GPT-6.1 SolOpenAI51.4%$0.62 an attempt
  4. Gemini 3.8 FlashGoogle51%$1.74 an attempt
  5. Claude Sonnet 5.5Anthropic41.9%$1.19 an attempt
  6. DeepSeek V4.1 FlashDeepSeek · Free to download37.1%$0.25 an attempt
  7. Kimi K3Moonshot AI · Free to download27.6%$2.55 an attempt
  8. GLM 5.3 FlashZ.ai · Free to download27.1%$0.24 an attempt
  9. Grok 4.7xAI18.1%$3.55 an attempt
  10. GPT-5.6 TerraOpenAI16.7%$1.17 an attempt

TreasuryQA Bench ↗A published study by Veris AI, which sells test copies of business software such as Dropbox and ranks no model of its own. Every agent reads the scans through the same document reading service, Reducto's, and each model runs in its maker's own agent, such as Claude Code or Codex, so another agent could change the scores. Board updated 28 Oct 2026 · Vailent checked 21 Nov 2026

Sort Requests and RFQsSend each emailed RFQ, order, or complaint to the right team.Jev 1.13 Specialist95.9%No prices publishedNone tested

A shared inbox is where orders wait. Sorting it right the first time is hours back every week, and a request nobody sees is a lost order.

A model built to decide, not to write, sorted 95.9% right with no training. Keyword rules got 77.2%. It is one small study, so treat it as a lead.

What to check

Always give it a “none of these” option: without one, it labelled a cake recipe as a technical question. Let it act only when it is sure, and send the rest to a person.

Read the highlightJev 1.13: A Model That Decides Instead of Writing

Tickets sorted into the right category.

  1. Jev 1.13Specialist95.9%
  2. Hand-written keyword rulesNo AI needed77.2%
  3. Word-count classifierNo AI needed66%

PriorBench ↗A one-person study on 400 made-up support tickets in French, not run by the model's maker. Board updated 20 Sep 2026 · Vailent checked 9 Nov 2026

Draft Replies Without Made-Up FactsDraft a customer reply from the order, the price, and the lead time.Finix S1 32B Ant Group · 32B1.8%No prices publishedLlama 3.3 70BMeta · 70BFree to download4.1%

A reply with an invented lead time or price is a promise your team has to break. The model that adds the fewest unsupported claims is the safest writer.

Small and open models add the fewest made-up claims, and the leader is a 32B model from Ant Group.

What to check

Give it only your approved facts, match every number, date, and promise against the order record, and have a person send.

Summaries with a claim the source does not support. Lower is better.

  1. Finix S1 32BAnt Group · 32B1.8%
  2. GPT-5.4 nanoOpenAI3.1%
  3. Gemini 2.5 Flash LiteGoogle3.3%
  4. Llama 3.3 70BMeta · 70B · Free to download4.1%
  5. DeepSeek V4.1 FlashDeepSeek · Free to download4.2%
  6. Gemma 3 12BGoogle · 12B · Free to download4.4%
  7. Qwen3 8BAlibaba · 8B · Free to download4.8%
  8. Granite 4.0 H SmallIBM · Free to download5.2%
  9. GPT-6 SolOpenAI6.5%
  10. Claude Haiku 4.5Anthropic9.8%
  11. Gemini 3 FlashGoogle13.5%

Vectara hallucination leaderboard ↗Run by Vectara, which grades every model with its own judge and ranks no model of its own. Board updated 22 Oct 2026 · Vailent checked 13 Nov 2026

Answer Service Requests From Company PolicyAnswer a customer's question on returns, credit terms, or delivery from your own policy documents.Qwen 3.8 Max Alibaba55.2%No prices publishedKimi K3Moonshot AIFree to download37.1%

A service reply that gets the policy wrong is a promise your team has to unwind. The model has to find the right document and then follow it.

The best model resolved about half of the requests on its first try, so a person still answers the rest. A custom system built around a model, Principled Machines' Cortex 0.1, resolved about three in four, verified by the board's maker on 3 November.

Kimi K3 and GLM-5.2 score the same, and both can run on your own servers. The board publishes no margin of error, so a gap of a few points may be chance. Its newest model came out in July 2026, and no model released since has a result.

What to check

Give it only your own policy documents, make it answer from them, and have a person approve anything that commits money or a date.

Banking customer-service requests resolved on the first try, from a library of about 700 policy documents.

  1. Qwen 3.8 MaxAlibaba55.2%
  2. Claude Opus 5Anthropic48.7%
  3. Grok 4.5xAI47.9%
  4. GPT-5.6 SolOpenAI46.9%
  5. GPT-5.5OpenAI44.6%
  6. Muse Spark 1.1Meta40.5%
  7. Kimi K3Moonshot AI · Free to download37.1%
  8. GLM-5.2Z.ai · Free to download37.1%

tau-bench Banking ↗Run by Sierra, which sells customer-service agents and ranks no model of its own. Model makers submit their own results, and the board gives no update date. The board states no update date · Vailent checked 9 Nov 2026

Research RegulationsFind what a rule requires before you ship a product to a new market.Gemini 4 Neon Google71.21%Muse EmberMeta54.55%103x cheaperKimi K3Moonshot AIFree to download63.64%

A missed requirement stops a shipment at the border. And the answer is only as good as the dated official text behind it.

The test cannot separate the three leaders from Muse Ember, which costs about 100 times less than Gemini 4 Neon a question and gives the most score for each dollar.

Google's Gemini 4 Neon scores 71.21% at $5.76 a question, and Gemini 4 Hydrogen and Claude Sonnet 5.5 both score 69.70%, at $4.84 and $11.67. Each score carries a margin of error of about 6 points on 66 questions of US administrative law, so the three share the top, and the test cannot separate them from 29 other models. Muse Ember, MiMo V2.6 Pro, GLM-5.3 Flash, Inkling 2 Small, and Muse Spark 1.3 Max all sit inside that margin, at under $0.60 a question. Kimi K3 is the best result you can run on your own servers, at $3.44 a question. GLM-5.3 is inside its margin, at $2.13.

What to check

Make it quote the dated official text, and check that text yourself. The test is US law, not REACH or GHS.

Read the highlightGPT-6.1 Sol: Near the Top for an Eighth of the BillRead the highlightClaude Sonnet 5.5: At the Top on Regulations, in a CrowdRead the highlightClaude Opus 5.5: First Place, and What It CostsRead the highlightClaude Haiku 5.5: Strong for Its Price, Weak on DocumentsRead the highlightMuse Wonder: The Top Version Only Runs in Muse Code

Regulatory research questions answered correctly.

  1. Gemini 4 NeonGoogle71.21%$5.76 a question
  2. Gemini 4 HydrogenGoogle69.7%$4.84 a question
  3. Claude Sonnet 5.5Anthropic69.7%$11.67 a question
  4. Gemini 4 ArgonGoogle68.18%$4.81 a question
  5. Claude Fable 5.1Anthropic68.18%$20.31 a question
  6. Claude Opus 5Anthropic65.15%$6.58 a question
  7. Claude Opus 5.5Anthropic65.15%$20.86 a question
  8. Gemini 4 CarbonGoogle65.15%$6.80 a question
  9. Kimi K3Moonshot AI · Free to download63.64%$3.44 a question
  10. Grok 4.6xAI62.12%$1.52 a question
  11. GLM-5.3Z.ai · Free to download62.12%$2.13 a question
  12. Muse Spark 1.3 MaxMeta60.61%$0.54 a question
  13. Inkling 2 SmallThinking Machines60.61%$0.24 a question
  14. GPT-5.6 SolOpenAI60.61%$19.69 a question
  15. Astra for Law 6.2Specialist60.61%$2.47 a question
  16. MiMo V2.6 ProXiaomi · Free to download59.09%$0.18 a question
  17. Claude Fable 5.2Anthropic59.09%$3.65 a question
  18. GPT-6.2 AstraOpenAI59.09%$11.82 a question
  19. GLM-5.3 FlashZ.ai · Free to download57.58%$0.20 a question
  20. Muse EmberMeta54.55%$0.056 a question
  21. GPT-6.1 SolOpenAI53.03%$2.76 a question
  22. DeepSeek V4.1 FlashDeepSeek · Free to download53.03%$0.25 a question
  23. GPT-6.2 SolOpenAI51.52%$3.54 a question
  24. Tencent Hy4 FlashTencent48.49%$0.055 a question
  25. MiMo V2.6 FlashXiaomi · Free to download46.97%$0.067 a question
  26. Inkling Mini 1117Thinking Machines21.21%$0.13 a question

vals.ai Legal Research ↗Run by vals.ai, an independent testing company. Board updated 19 Nov 2026 · Vailent checked 21 Nov 2026

Find a Clause in a ContractFind the price-adjustment or force majeure clause in a supply agreement.Claude Opus 5 Anthropic73.19%DeepSeek V4 Pro 0813DeepSeek65.42%394x cheaperKimi K3Moonshot AIFree to download71.56%

The clause you cannot find is the one that costs you in a dispute. Page citations let a person check the answer in seconds.

DeepSeek V4 Pro 0813 gives the most score for the money: 7.8 points below the leader at about $0.0022 a question. The test reused stored input for almost all its input tokens, which lowers the bill. The test has not run the newest models.

What to check

Require a page citation for every answer, open it, and treat “not found” as a result to check.

Questions on long credit agreements answered correctly.

  1. Claude Opus 5Anthropic73.19%$0.86 a question
  2. Claude Fable 5Anthropic71.83%$1.74 a question
  3. Kimi K3Moonshot AI · Free to download71.56%$0.064 a question
  4. Muse Spark 1.1Meta71.29%$0.11 a question
  5. Inkling SmallThinking Machines · Free to download69.62%$0.025 a question
  6. Grok 4.3xAI68.53%$0.13 a question
  7. GPT-5.5OpenAI68.42%$0.43 a question
  8. DeepSeek V4 Pro 0813DeepSeek65.42%$0.002 a question

vals.ai CorpFin v2 ↗Run by vals.ai, an independent testing company. Board updated 12 Aug 2026 · Vailent checked 21 Nov 2026

Search Years of EmailFind what was promised to a customer three years ago.Keyword search (BM25) No AI needed87.5%No prices publishedNone tested

Commitments hide in old threads. The fastest way to find them is not always the newest technology.

Plain keyword search found the right email far more often than a neural search model. Use AI to read what it finds.

What to check

The study's questions name specific people and things, which suits keyword search. Look for later messages that changed the commitment.

Right email found in the top five results.

  1. Keyword search (BM25)No AI needed87.5%
  2. ColBERTv2Specialist · Free to download54.1%

EnronQA ↗A published study on 103,638 real business emails. Board updated 1 May 2025 · Vailent checked 9 Nov 2026

Write a Report From Several SourcesTurn sales data, notes, and market news into a quarterly account review.Claude Opus 5.5 Anthropic1866No prices publishedKimi K3.1 ProMoonshot AIFree to download1861

Some work has to be right and look right for a customer or a board. Here a large model earns its price.

For finished professional work, large models lead, Sonnet 5.5 matches Opus 5.5 at half the price, and the open Kimi K3.1 Pro is level with both.

Opus 5.5, Kimi K3.1 Pro, Muse Wonder Full, Sonnet 5.5, and Anthropic's new Claude Fable 5.2 are level within the test's margin of error. Sonnet and Kimi K3.1 Pro cost half as much per token as Opus. Meta offers Muse Wonder Full only inside Muse Code, its coding tool, not through its API. Moonshot AI put Kimi K3.1 Pro's weights on Hugging Face on 2 November, so it is now the best result you can run on your own servers, and its license puts no conditions on a company's internal use. It is a very large model, with about 2.8 trillion parameters, so running it yourself takes a large cluster of servers. DeepSeek V4.1 Flash 1017 is the next open model, 115 points back. Google's Gemini 4 Carbon rates 1733, and since 10 November any paying business can use it.

What to check

Check every figure against its source before it leaves the building.

Read the highlightGPT-6.1 Sol: Near the Top for an Eighth of the BillRead the highlightClaude Sonnet 5.5: At the Top on Regulations, in a CrowdRead the highlightClaude Opus 5.5: First Place, and What It CostsRead the highlightClaude Haiku 5.5: Strong for Its Price, Weak on DocumentsRead the highlightMuse Wonder: The Top Version Only Runs in Muse Code

Preference rating on real professional tasks.

  1. Claude Opus 5.5Anthropic1866
  2. Kimi K3.1 ProMoonshot AI · Free to download1861
  3. Muse Wonder FullMeta1857
  4. Claude Sonnet 5.5Anthropic1838
  5. Claude Fable 5.2Anthropic1831
  6. Qwen4 MaxAlibaba1809
  7. Muse WonderMeta1770
  8. Claude Fable 5.1Anthropic1765
  9. DeepSeek V4.1 Flash 1017DeepSeek · Free to download1746
  10. Gemini 4 CarbonGoogle1733
  11. Claude Opus 5Anthropic1729
  12. Grok 4.7xAI1718
  13. MiMo V2.6 ProXiaomi · Free to download1692
  14. GLM-5.3Z.ai · Free to download1655
  15. GPT-6.1 SolOpenAI1607

Artificial Analysis GDPval-AA ↗Run by Artificial Analysis, an independent testing company. The board states no update date · Vailent checked 16 Nov 2026

Find the Current Answer When Records DisagreeFind which price applies when a later email revises the one in the PDF quote.GPT-5.6 Sol OpenAI68%No prices publishedQwen3.5-27BAlibaba · 27BFree to download50.3%

A price, a spec, or a date changes in a later email, chat, or revised file. An answer built on the old record quotes the wrong number.

The best model answered about two questions in three, and on the questions where two records disagree it got only 55 to 56% right.

What to check

Make it name the date and source of every fact, and check the newest record yourself. The study tested no Claude model and published no prices.

Questions answered correctly, with 50 records found for each.

  1. GPT-5.6 SolOpenAI68%
  2. GPT-5.6 LunaOpenAI65.8%
  3. Gemini 3.8 FlashGoogle63.9%
  4. Gemini 3.1 ProGoogle62.5%
  5. Qwen3.5-27BAlibaba · 27B · Free to download50.3%
  6. Qwen3.5-4BAlibaba · 4B · Free to download43.8%
  7. Qwen3.5-9BAlibaba · 9B · Free to download43.1%

MMEnterprise-Bench ↗A published study by KAIST and Korea University, on 1,613 questions over work records in chats, notes, emails, and PDFs that change over time. Board updated 20 Oct 2026 · Vailent checked 9 Nov 2026

Match Products and GradesMatch a customer's product name to the grade in your catalogue.Kimi K2.6 Moonshot AIFree to download87.2No prices publishedSame as the best resultFree to download and run in-house

The same grade has five names across suppliers and customers. A wrong match ships the wrong material.

An open model from Moonshot matched products best, ahead of GPT-5.2.

What to check

Match exact identifiers first. Send every unsure match, and any grade equivalence that changes a price, to an expert.

Product matches made correctly (F1), out of 100.

  1. Kimi K2.6Moonshot AI · Free to download87.2
  2. GPT-5.2OpenAI84.47
  3. DittoSpecialist · Free to download71.23

Steiner and Bizer, WDC Products ↗A published study by the University of Mannheim. Board updated 5 Oct 2026 · Vailent checked 9 Nov 2026

Forecast Demand for Slow-Moving StockPlan stock for a grade that sells a few drums a quarter.ADIDA No AI needed90.6%No prices publishedSame as the best resultFree to download and run in-house

Slow, lumpy demand is where stock-outs and dead stock both come from. The right method trades service against the cash tied up in stock.

Classic forecasting methods filled far more complete orders than AI forecasting models, at the cost of more stock on the shelf.

The classic methods fill more orders because they forecast high, so they hold more stock: Croston carried about four times the inventory of TimesFM. When each method's safety stock is tuned, the gap narrows.

What to check

Backtest on your own order history, and compare fill rate and stock held together, never one without the other.

Complete orders filled from stock.

  1. ADIDANo AI needed90.6%
  2. Croston's methodNo AI needed85.8%
  3. SBANo AI needed84.7%
  4. LightGBMNo AI needed77.1%
  5. MoiraiSpecialist · Free to download71.9%
  6. Chronos-Bolt smallSpecialist · Free to download56.7%
  7. TimesFMSpecialist · Free to download54.4%

SMU intermittent demand study ↗A published study by Singapore Management University and ST Logistics, on 20,330 real orders for industrial spare parts. Board updated 12 Sep 2026 · Vailent checked 9 Nov 2026

For products that sell steadily, AI forecasting models do better. Several are small and free. The best licensed for business use is t0-gamma, under Apache 2.0. TimeH2O-v1 and Timestral-209M score higher, but neither is licensed for business use, and nor is TimesFM-3.

  1. TimeH2O-v1Specialist91.5%
  2. Timestral-209MSpecialist · 209M89%
  3. t0-gammaSpecialist · Free to download87.4%
  4. TimesFM-3Specialist · 331M82%
  5. iAmTimeSpecialist · Free to download76.4%
  6. Chronos-2Specialist · 119M · Free to download · The test's own product75.8%
  7. t0-betaSpecialist · 256M · Free to download73.4%
  8. TiRex-2Specialist · 38M · Free to download72.4%
  9. Toto-2.0 2.5BSpecialist · 2.5B · Free to download72.2%
  10. AutoARIMANo AI needed28%

fev-bench ↗Run by the AutoGluon team at Amazon, which also makes Chronos, a model it ranks. Board updated 28 Oct 2026 · Vailent checked 2 Nov 2026

Build and Check Spreadsheet ModelsBuild a margin model in Excel from your price list and cost sheet, then check its formulas.Claude Fable 5.2 Anthropic77.38%Tencent Hy4 FlashTencent62.97%196x cheaperKimi K3Moonshot AIFree to download71.26%

One wrong formula gives a confident wrong number. A model that builds the sheet saves hours only if its logic holds up.

The best models build these spreadsheets right about three times in four. Muse Ember is inside the leader's margin at about a seventieth of the price, and Tencent Hy4 Flash costs about a two hundredth but gets fewer right.

The test cannot separate Claude Fable 5.2 from 19 other models. Google's new Gemini 4 Neon scores 76.46% at $3.82 a task and OpenAI's new GPT-6.2 Astra 76.03% at $2.80, both well under Fable 5.2's $6.67, and GPT-6.2 Sol is inside the margin too, at $0.84. Claude Fable 5.1 costs $15.93 a task, more than twice the leader's price. Tencent Hy4 Flash gives the most score for each dollar, 62.97% at $0.034 a task, but it sits outside the leader's margin. Muse Ember is the cheapest model inside it, at $0.09 a task. Kimi K3 is the best result you can run on your own servers, and MiMo V2.6 Flash, Mistral Large 4, and Kimi K3.1 Pro, all open, are inside its margin.

What to check

Check every formula and every input cell against your own figures before a number leaves the sheet. The test is financial models, not your price lists.

Excel financial modeling tasks graded correct.

  1. Claude Fable 5.2Anthropic77.38%$6.67 a task
  2. Claude Fable 5.1Anthropic76.73%$15.93 a task
  3. Gemini 4 CarbonGoogle76.47%$5.76 a task
  4. Gemini 4 NeonGoogle76.46%$3.82 a task
  5. GPT-6.2 AstraOpenAI76.03%$2.80 a task
  6. Claude Sonnet 5.5Anthropic75.9%$6.53 a task
  7. Claude Opus 5.5Anthropic75.84%$9.33 a task
  8. Gemini 4 HydrogenGoogle75.8%$3.17 a task
  9. Muse Wonder FullMeta75%$4.58 a task
  10. Gemini 4 ArgonGoogle74.83%$11.80 a task
  11. GPT-6.2 SolOpenAI74.29%$0.84 a task
  12. GPT-6.1 TerraOpenAI73.33%$0.56 a task
  13. Muse EmberMeta72.17%$0.091 a task
  14. Kimi K3Moonshot AI · Free to download71.26%$3.07 a task
  15. GPT-6.1 SolOpenAI71.09%$1.14 a task
  16. GPT-6 LunaOpenAI69.95%$0.14 a task
  17. MiMo V2.6 FlashXiaomi · Free to download68.88%$0.11 a task
  18. Mistral Large 4 OSSMistral AI · Free to download68.67%$0.83 a task
  19. MiMo V2.6 ProXiaomi · Free to download67.35%$0.33 a task
  20. Inkling Mini 1117Thinking Machines66.02%$1.01 a task
  21. Claude Haiku 5.5Anthropic65.48%$1.36 a task
  22. Tencent Hy4 FlashTencent62.97%$0.034 a task

vals.ai Excel Modeling Benchmark v1.1 ↗Run by vals.ai, an independent testing company. Its tasks are financial models such as buyouts and cash-flow valuations, not price lists. Board updated 19 Nov 2026 · Vailent checked 21 Nov 2026

Screen Sanctions ListsCheck a new customer or its owners against sanctions lists.GPT-4o OpenAI98.95No prices publishedDeepSeek R1 Distill Qwen 14BDeepSeek · 14BFree to download98.23

One missed match is a regulatory breach, and every false alarm is a person's time. The score below weighs both.

An open 14B model comes within a point of the best. The production matcher misses almost nothing, but one pair in seven it flags is a false alarm.

What to check

A person makes every decision. Check every hit, every name in another script, and a sample of non-hits.

Matches judged correctly, balancing misses and false alarms (F1), out of 100.

  1. GPT-4oOpenAI98.95
  2. GPT-5.2 ProOpenAI98.75
  3. DeepSeek R1 Distill Qwen 14BDeepSeek · 14B · Free to download98.23
  4. Claude 3.7 SonnetAnthropic97.5
  5. Claude Opus 4.5Anthropic95.45
  6. Llama 3.1 8BMeta · 8B · Free to download94.05
  7. Feature-based matcherNo AI needed91.33

OpenSanctions Pairs ↗A published study on OpenSanctions data. Board updated 25 Aug 2026 · Vailent checked 9 Nov 2026

Model highlights and updates.

New models and new test results, dated, and the tasks they change.

  1. Last Week in AI for Chemicals

    OpenAI's GPT-6.2 Astra and Google's Gemini 4 Neon arrived a day apart, and Neon now shares first place on regulatory research.

    Read the edition
  2. Last Week in AI for Chemicals

    Claude Fable 5.2 now builds spreadsheet models best and leads the general index, and two cheaper models joined the top group on regulatory research.

    Read the edition
  3. Last Week in AI for Chemicals

    Kimi K3.1 Pro, level with the best model on real professional work, is now free to run on your own servers.

    Read the edition
  4. Last Week in AI for Chemicals

    Google's Gemini 4 Carbon ties for first on vals.ai's legal research test and is not yet open to businesses.

    Read the edition
  5. Last Week in AI for Chemicals

    DeepSeek's V4.1 Flash update ranks 7th of 27 on real professional work, the best result a company can run on its own servers.

    Read the edition
  6. Last Week in AI for Chemicals

    Meta's Muse Spark Open ties Claude Haiku 5.5 on US regulatory research and can run on your own servers.

    Read the edition
  7. Muse Wonder Released

    Meta's most capable model ranks 7th of 29 on rated professional work, at Claude Opus 5.5's token price. Muse Wonder Full ranks 3rd, level with Opus, and runs only inside Meta's Muse Code.

    Read the highlight
  8. Last Week in AI for Chemicals

    Meta's Muse Wonder ranks 4th of 27 on real professional work at Claude Opus 5.5's token price, and the version level with Opus runs only inside Muse Code, so a company can't build it into its own systems.

    Read the edition
  9. Claude Haiku 5.5 Released

    Level with Inkling 2 Small, GPT-6.1 Sol, GLM-5.3, and Ling 3.1 Flash 1019 on rated professional work, at $0.41 an index task. It is 37th of 95 on regulatory research and 104th of 157 on reading whole documents.

    Read the highlight
  10. Last Week in AI for Chemicals

    Kimi K3, an open model you can run on your own servers, now leads the open models on regulatory research.

    Read the edition
  11. GPT-6.1 Sol Arrives

    It scores 54.6 on the general intelligence index, 5.4 points behind the leader, at about an eighth of Claude Opus 5.5's cost a task. On regulatory research it scores 53% at $2.76 a question, just inside the leading group.

    Read the highlight
  12. Sonnet 5.5 Ties for First on Regulatory Research and Joins the Document Test

    Tied first of 95 models on regulatory questions with Gemini 4 Hydrogen, and the test cannot separate it from a model at about a two-hundredth of the price. On tables it lands 16th, at 2.74 cents a page.

    Read the highlight
  13. Claude Opus 5.5 Released

    First on the priced table test at its high setting, and first on finished professional work, level with Sonnet 5.5, whose tokens cost half as much. It is the dearest model in the regulatory leading group.

    Read the highlight
  14. Jev 1.13 Released

    A model that returns a decision and how sure it is, not text. In one small study it sorted 95.9% of tickets right with no training.

    Read the highlight
  15. A Free 1.2B Model Tops the Independent Document Test

    TeleOCR took first place on OmniDocBench on whole documents and on tables, ahead of Gemini 3 Pro. Since 29 October it is level with OXR-1.0 on both, and an outside tester challenged its table scores on 26 October.

    Read the highlight

Why one model is not enough.

One model was never the answer.

For years, using AI meant picking one model and asking it everything. The table above shows why that never held. A small specialist reads your documents best, classic forecasting fills more orders for your slow-moving stock, and the most expensive model is rarely the right buy.

Information is not context.

A model knows the public internet. It does not know your price lists, your allocation rules, your customers' history, or the exception your senior trader remembers. Give a smaller model that context, and it does work a larger one without it gets wrong.

The right size for every step.

Sorting a request, reading a certificate, and drafting a reply to a key account are different work. Small models should carry the volume, the largest should be saved for the work that needs them, and every answer should come back with its sources, ready for your review.

No two chemical companies run the same way, so the AI that serves them should not either.

See how Vailent works