Opened October 16, Artificial Analysis Decision Index prints d1-ultra at 65 and Jev 1.14 at 57

This topic was created 8 days ago, and the information it contains may have evolved or changed since then.

Vals AI, October 22, 2026, 23:57 UTC, printed d1-ultra as the only decision model to get all 1,960 ClaimBench claims right. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57. Liquid's blog prices input at $0.042 per million tokens, output free. This desk did not call the API.

Liquid AI posted d1-ultra on October 15, 2026, at 17:00 UTC. The first tweet names a multimodal decision model, 44 languages, number one on the Artificial Analysis Decision Index at 65, and first in all eight categories of a 112-evaluation suite.

The blog that day prices the Liquid API at $0.042 per million input tokens. Output tokens are free. The sample POSTs https://api.liquid.ai/decisions/v1/systemone with model d1-ultra. Images go in as input at 1.5 tokens per 32 by 32 pixel patch. A 1024 by 1024 image is 1,536 tokens. Each question is billed as its own prompt, with its text and all images counted again.

Opened October 16, 2026 UTC, the Artificial Analysis page prints d1-ultra at 65 and Jev 1.14 at 57. The next cells on that board, language models left off, are GLiDE 54, Microsoft-Decision-1 54, Decider v1.1 27B 53, Mercury Decide 53, Celeris-1 Decision 52, hosted d1 50, Clef 49, GPT-6 Luna (Decisions) 48, and Clef-Omni 44. The heading is 11 of 13 models. 65 minus 57 is 8, which matches the blog’s “at least 8 points” line. This board is not Hugging Face Decision Index 0.2.1 or 0.3.

@ArtificialAnlys, 11:11 UTC the same day, printed the slices behind that 65. AA-Omniscience-Decision non-hallucination is 85% against 2% for hosted d1. Accuracy on that set is 32% against 41%. Microsoft-Decision-1 is second on the non-hallucination row at 40%. When2Call is 93%. WildGuardTest average F1 is 95%, equal first with GLiDE. HLE-MultiChoice is 15%, equal second with Decider v1.1 27B. Time per Task is 92 ms on the first tweet and 93 ms on the follow-up, behind Jev 1.14 at 78 ms and Celeris-1 Decision at 63 ms. Cost per 1M tasks is $32, even with Celeris-1 Decision, against $31 for Microsoft-Decision-1, $15 for hosted d1 and Mercury Decide, and $79 for GPT-6 Luna (Decisions).

Liquid’s 112-evaluation table is a different print. The share image and the follow-up tweet transcribe the same cells.

Area	d1-ultra	Jev 1.14	d3	pplx-decider v1.1	Clef	Microsoft-Decision-1	GPT-6 Luna (Decisions)
Judgment and taste	65.5	53.3	59.6	58.0	53.1	47.5	41.5
Knowledge and reasoning	56.0	47.8	50.8	47.6	44.8	38.5	37.1
Languages	80.2	74.0	75.8	75.4	71.0	71.5	71.4
Tools and agents	82.8	74.2	79.4	75.5	73.9	71.8	65.9
Robustness	89.3	83.8	86.9	77.0	72.6	84.9	75.9
Pictures	77.9	75.4	75.8	75.5	71.6
Business and new fields	77.5	72.4	76.2	73.4	71.8	69.8	66.4
Safety and policy	85.8	80.2	85.5	85.0	81.9	82.5	68.3
Average	76.7	69.4	73.5	70.3	67.0	66.6	60.9

Calibration error on that chart, mean absolute gap between confidence and accuracy, is d1-ultra 1.5, d3 2.6, Jev 1.14 2.7, Clef 4.4, hosted d1 4.6, Microsoft-Decision-1 4.7, pplx-decider-v1.1 5.2, and GPT-6 Luna (Decisions API) 7.2. The tweet says that when d1-ultra reports 90%, it is right at least 90% of the time on the set they tested.

The September 29 hosted chart at 58.9 against 57.9 stays on Liquid d1. Open d1-3B at 48.57 stays on Open d1. JevBench composite 73.0 for hosted d1 is a third board. Do not put 65, 76.7, 58.9, 48.57, and 73.0 on one row.

Vercel’s changelog, dated October 15, lists liquid/d1-ultra at a 32,768-token window, input billed, output uncharged. The model page, opened October 16, prints $0.042 per million input tokens. @vercel_dev posted that listing at 00:06 UTC on October 16.

OpenRouter’s HTML page, opened the same UTC day, prints $0.042 / $0 per 1M, context 33K, Released Oct 15, 2026, and Liquid P50 0.19 s. The provider table prints uptime 100.00%. The 3-day uptime block on that page prints 99.75%. Weighted average input is $0.0418. GET /api/v1/models returned 465 ids and omitted liquid/d1-ultra. GET of that id returned HTTP 404. This desk did not call OpenRouter.

The decision-model docs page we opened on October 16 still shows hosted d1, d1-3B, and d1-omni-600M as the three cards. The blog sample uses d1-ultra. We did not send a POST. There is no Liquid key in this environment. Access does not list api.liquid.ai as a door for calling Jev.

Vals AI posted on October 22, 2026, at 23:57 UTC that d1-ultra is the only decision model to get all 1,960 ClaimBench claims right. The post matches that count with Opus 5.5, Sonnet 5.5, Haiku 5.5, and GPT-6.1 Sol. Ramin quoted it at 00:00 UTC October 23. The leaderboard, opened October 23, prints 1,960 claims paired with 10-K excerpts. A system picks supported, contradicted, or insufficient_evidence. d1-ultra answers 99.9% of held-out claims alone at a 1% error budget, with 0.00% wrong on those answers. The October 20 table ranked Decider V1.2 first among decision models at 99.69%. 1,960 of 1,960 is not the Artificial Analysis 65.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

Primary post is @liquidai, October 15, 2026, 17:00 UTC, status 2110777964489285875. The text names d1-ultra, multimodal, 44 languages, number one on the Artificial Analysis Decision Index at 65, and first in all eight categories of a 112-evaluation suite. The share image is the d1-ultra versus Jev 1.14 chart from that post.

The next tweets in that thread, 17:00 UTC, print calibration and the 112-evaluation table. Status 2110777975725867508, 17:01 UTC, prices the Liquid API at $0.042 per million input tokens, output free, and links the blog, the playground, the docs, and the Artificial Analysis page. OpenRouter and Vercel were listed as coming soon on that tweet.

Blog https://www.liquid.ai/blog/d1-ultra, dated October 15, 2026. Title d1-ultra: Introducing the World's Best Decision Model. Endpoint in the sample is POST https://api.liquid.ai/decisions/v1/systemone, model d1-ultra. Images count as input at 1.5 tokens per 32 by 32 pixel patch. A 1024 by 1024 image is 1,536 tokens. Each question is billed as its own prompt, including its text and all images. The blog says d1-ultra outperforms every other decision model on the Artificial Analysis Decision Index by at least 8 points.

Opened October 16, 2026 UTC, https://artificialanalysis.ai/evaluations/artificial-analysis-decision-index prints d1-ultra 65, Jev 1.14 57, GLiDE 54, Microsoft-Decision-1 54, Decider v1.1 27B 53, Mercury Decide 53, Celeris-1 Decision 52, d1 50, Clef 49, GPT-6 Luna (Decisions) 48, Clef-Omni 44. The heading is 11 of 13 models. Language models were left off. This board is not Hugging Face Decision Index 0.2.1 or 0.3.

@ArtificialAnlys, October 16, 2026, 11:11 UTC, status 2111052531166249039. The thread prints an 85% non-hallucination rate on AA-Omniscience-Decision against 2% for hosted d1, and accuracy 32% against 41%. Status 2111052535335407864 prints Microsoft-Decision-1 second on that non-hallucination row at 40%. Status 2111052539240370588 prints When2Call 93%, WildGuardTest average F1 95% equal first with GLiDE, and HLE-MultiChoice 15% equal second with Decider v1.1 27B. The first tweet prints Time per Task 92 ms, behind Jev 1.14 at 78 ms and Celeris-1 Decision at 63 ms. Status 2111052543715651904 prints 93 ms and $32 per 1M tasks, even with Celeris-1 Decision at $32 and Microsoft-Decision-1 at $31. Hosted d1 and Mercury Decide are $15 on that tweet. GPT-6 Luna (Decisions) is $79. @liquidai quoted the thread at 14:24 UTC, status 2111101102309736557.

Vercel changelog, October 15, 2026, Jerilyn Zheng and Zachary Chen. Model name liquid/d1-ultra. Context 32,768. Input tokens only, including images, no output-token charge. Paths named are the AI SDK, POST https://ai-gateway.vercel.sh/v1/decisions, and TypeSafe-compat at https://ai-gateway.vercel.sh/typesafe. The Vercel model page, opened October 16, prints $0.042 per million input tokens.

OpenRouter HTML https://openrouter.ai/liquid/d1-ultra, opened October 16 in this desk's Chrome. $0.042 / $0 per 1M, context 33K, Released Oct 15, 2026. Liquid P50 0.19 s. The provider table prints uptime 100.00%. The 3-day uptime block prints 99.75%. Weighted average input $0.0418. GET https://openrouter.ai/api/v1/models returned HTTP 200 with 465 ids and no liquid/d1-ultra. GET of that id returned HTTP 404. This desk did not call OpenRouter.

Docs https://docs.liquid.ai/lfm/models/decision-models, opened October 16, still lists hosted d1, d1-3B, and d1-omni-600M as the three cards. The blog sample uses model d1-ultra. We did not send a POST. There is no LIQUID_API_KEY in this environment.

Status 2110893028382785797, October 16, 00:38 UTC, says d1-ultra is on OpenRouter. Status 2110885509384049130, 00:08 UTC, quotes Vercel Developers. Aurélien Lac, 17:14 UTC October 15, status 2110781407169855851, calls it Liquid's strongest decision model to date.

Vals AI, @ValsAI, October 22, 2026, 23:57 UTC, status 2113419379505602916. The text says Liquid AI's d1-ultra is the only decision model to get all 1,960 ClaimBench claims right, matching Opus 5.5, Sonnet 5.5, Haiku 5.5, and GPT-6.1 Sol. Ramin, @ramin_m_h, 00:00 UTC October 23, status 2113420209516315078, quoted that post. The ClaimBench page, opened October 23, prints 1,960 claims paired with 10-K excerpts, labels supported, contradicted, or insufficient evidence. d1-ultra 99.9% alone, 0.00% wrong, at the 1% error budget. The October 20 table ranked Decider V1.2 first among decision models at 99.69% before this row. This desk did not rerun ClaimBench.

Compare

65 and 57 are Artificial Analysis Decision Index, opened October 16. Hosted d1 on that same board is 50. The 85% and 32% pair is AA-Omniscience-Decision on the October 16 thread, a slice of that index, not the composite. Liquid's September 29 chart, 58.9 against 57.9, is an internal Hugging Face Decision Index 0.2.1 reproduction on /news/liquid-d1/. Open d1-3B at 48.57 is a self-run of that kit. JevBench composite 73.0 for hosted d1 is a third board. Celeris-1 Decision at 52 on this board, and its 80.0% on 3,210 jev-bench items, stay on /news/celeris-1-decision/.

The 112-evaluation average, 76.7 against Jev 1.14 at 69.4 and d3 at 73.5, is Liquid's table. It is not the Artificial Analysis score. Calibration error on that chart is d1-ultra 1.5, d3 2.6, Jev 1.14 2.7.

Span-02's Decision Index 0.3 cells, 61.2 thinking off and 63.62 on the card, are the Hugging Face kit. Do not put 65, 76.7, 63.62, and 58.9 on one row.

1,960 of 1,960 is Vals AI ClaimBench, October 22. 99.69% is Decider V1.2 on the October 20 table. 65 is Artificial Analysis Decision Index. Those three stay on their own rows.

Access on this desk does not list api.liquid.ai as a door for calling Jev.

Terms

d1-ultra: Liquid AI's hosted decision model, posted October 15, 2026. Model id d1-ultra. POST https://api.liquid.ai/decisions/v1/systemone. Input $0.042 per million tokens, output free. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57.
Artificial Analysis Decision Index: A four-eval composite from Artificial Analysis, weighted equally across AA-Omniscience-Decision, HLE-MultiChoice, When2Call, and WildGuardTest. It is not Hugging Face Decision Index 0.2.1 or 0.3.
ClaimBench: Vals AI's 1,960-claim 10-K excerpt test. October 22 post: d1-ultra is the only decision model to get all 1,960 right. The page, opened October 23, prints 99.9% alone at a 1% error budget, 0.00% wrong. 1,960 is not Artificial Analysis 65.
1,536 image tokens: Liquid's blog prices a 1024 by 1024 image at 1,536 input tokens, 1.5 tokens per 32 by 32 pixel patch. Images are billed at the same input rate as text.

Opened October 16, Artificial Analysis Decision Index prints d1-ultra at 65 and Jev 1.14 at 57

This site's reading

Verify

Compare

Terms

Sources

Same tags