Published
Updated
SubQ Snap 1.1 prints 65 on Artificial Analysis Decision Index, $0.04 per million input
This topic was created 8 days ago, and the information it contains may have evolved or changed since then.
On October 23, 2026, Date headers 02:23:17 GMT to 02:23:20 GMT, three POSTs from this desk to SubQ Snap returned HTTP 200. Docs name subq-snap-1.1 at $0.04 per million input tokens. GET /v1/models listed subq-snap-1.1 and subq-snap-preview. Artificial Analysis Decision Index prints Snap 1.1 at 65, tied with d1-ultra and Elephant 9B. The October 12 blog prints Decision Index 0.2.1 at 61.7 against a Jev cell of 57.9.
Alexander Whedon posted SubQ Snap on October 12, 2026, at 16:03 UTC. The bio on that post reads Co-Founder, CEO at @subquadratic. The visit link resolved to subq.ai. The attachment is a video of about 76 seconds. We did not transcribe a score from it. The share image on this article is that video’s still.
The October 12 blog calls Snap a decision model for agent workflows. It says the model reads up to 384K tokens of text and images, finishes a request of up to 100K tokens in about one second, and sits four points above Jev on Decision Index v0.2.1. The post says compute per token stays constant as context grows. Those lines are Subquadratic’s.
Docs name the public-preview id subq-snap-preview. Input is $0.05 per million tokens. Output is free. The path is POST https://api.subq.ai/v1/systemone. The same page lists subq-preview for long-context chat, status Early access. The quickstart signs up at console.subq.ai and stores the key as SUBQ_API_KEY. A request without credits, it says, returns 402.
Opened October 13, 2026 UTC, GET https://api.subq.ai/v1/models with no key returned HTTP 200. The list had one id, subq-snap-preview, owned_by subquadratic. It did not list subq-preview. An empty POST to /v1/systemone returned HTTP 422, missing state. A POST that included a state and a noul, still with no key, returned HTTP 401, missing API key. The same body with Authorization: Bearer sk-test returned HTTP 401, missing or invalid API key. We did not create a console account.
On October 13, 2026, at about 03:00 UTC, GET /v1/models with a bearer key kept in the environment returned HTTP 200. The list was the same one id, subq-snap-preview, name SubQ Snap Preview, owned_by subquadratic, created 0. It still did not list subq-preview. Content-Length was 135 with the key and without it.
Two POSTs to https://api.subq.ai/v1/systemone with model subq-snap-preview returned HTTP 200. The state was the Stripe sentence this desk already used on October 6 with Jev and with OpenAI Decisions: “Hi, I’ve been trying to connect my Stripe account for 3 days and the integration keeps failing. I’m losing sales. Please help ASAP.” This page does not print the request JSON. No image went out. The noul questions included true and false criteria, which the SubQ docs example includes.
Urgency alone, instruction “Does this message express urgency?”, returned noul 0.98. Usage was 119 input tokens and 0 output tokens. The response had no processing-time header. The clock on this desk, from send to the full body, was 1,195 ms.
The three-question call chose technical, confidence 0.90, with probabilities 0.93 on technical, 0.06 on billing, and 0.01 on sales. Frustration scored 0.98, confidence 0.93, with 0.95 of the weight on “Frustrated but civil,” 0.04 on “Calm, just stating facts,” and 0.01 on “Very angry, strong language.” The urgency noul was 0.98. Usage was 235 input tokens and 0 output tokens. The clock was 2,291 ms. Both responses named model subq-snap-preview. usage.output_tokens was 0. At $0.05 per million input tokens, 119 tokens is $0.00000595 and 235 tokens is $0.00001175. Neither body included an invoice line.
The Decisions contract is the System One shape. state is the text, object, or array to judge. questions are named noul, choice, or score items. A noul is a yes/no probability from 0 to 1. A choice has 2 to 255 named options and returns choice, a probability for every option, and confidence. A score has 2 to 10 ordered levels and returns a probability-weighted score that can fall between levels. Several questions share one request. usage.output_tokens is always 0.
The window is 384K tokens, printed as 393,216. A request over that limit, or close to it, fails with 400. Images are optional: up to four, PNG, JPEG, or WebP, 4 MiB decoded each, 8 MiB decoded in total, about 64 to 256 image tokens each. The request body can be 13 MiB. Video is not supported. The docs say to sample up to four frames and send them as images. The October 12 post calls 384K twelve times Jev. TypeSafe’s models page lists 32k for state plus the longest question and 64k for the whole request.
The blog’s Decision Index paragraph scores Snap at 61.70 on edition 0.2.1, against 57.90 for Jev. The table and the homepage round those cells to 61.7 and 57.9. Every request in that suite, the blog says, is under about 34K tokens. Snap leads in four of five areas: arts and human taste +12.11, language understanding +7.46, retrieval and classification +6.27, and tools and automation +3.39. The fifth area is unnamed on that page.
A second table takes 16 Decision Index tasks the blog maps to enterprise work. The average row is 67.54 against 61.6. The TL;DR rounds Snap to 67.6. The homepage chart prints 67.5 versus 61.6. Snap leads on 13 of the 16 rows and trails on three: ACOS −0.6, API-Bank −1.6, When2Call −7.6. The largest printed gain is PhishNChips, 47.1 against 25.1.
| Workflow | Benchmark | Snap | Jev | Gain |
|---|---|---|---|---|
| Support routing | CLINC150+OOS | 91.5 | 89.2 | +2.3 |
| Support routing | BANKING77 | 80.6 | 79.5 | +1.1 |
| Security | PhishNChips | 47.1 | 25.1 | +22.0 |
| Search | Amazon ESCI | 48.7 | 43.8 | +4.9 |
| Search | BRIGHT | 44.3 | 40.6 | +3.7 |
| Document checks | RAGTruth | 64.8 | 51.3 | +13.5 |
| Document checks | HoVer | 54.2 | 45.7 | +8.5 |
| Legal | ContractNLI | 70.7 | 59.1 | +11.6 |
| Finance | FinEntity | 92.8 | 80.8 | +12.0 |
| Clinical | NLI4CT | 72.8 | 69.0 | +3.8 |
| Code | CRUXEval | 73.8 | 57.1 | +16.7 |
| Tools | ToolRet | 63.1 | 59.9 | +3.2 |
| Tools | BFCL | 96.2 | 94.3 | +1.9 |
| Tools | API-Bank | 86.4 | 88.0 | −1.6 |
| Tools | When2Call | 67.0 | 74.6 | −7.6 |
| Reviews | ACOS | 26.7 | 27.3 | −0.6 |
| Average (16) | 67.54 | 61.6 | +5.96 |
Bespoke Nimble on that blog is 82.9% against 76.0% for Jev, 3,880 human-labelled examples from 13 public datasets.
Decision RULER, the blog says, turns 11 of RULER’s 13 tasks into four-option decisions and drops the two question-answering tasks. The printed table stops at 256K. The body and the teaser still extend the run to 1M tokens.
| Context | Snap | Jev |
|---|---|---|
| 32K | 99.8 | 94.1 |
| 64K | 99.8 | N/A |
| 128K | 99.5 | N/A |
| 256K | 99.8 | N/A |
The teaser alt lists 99.3 and 96.3 after those four cells, from 32K through 1M. The TL;DR names 96.3 at 1 million tokens.
Decision LCR converts Artificial Analysis AA-LCR. 100 questions over real multi-document sets of 72K to 115K tokens become 118 decisions. Each decision runs three times with shuffled options. A question counts only when every part is right. Chance is 19.2. With the documents removed, models score 24 to 25. Snap, reading the documents in one shot, is 74.0. Jev, on chunked input, is 48.0. The blog says those scores cannot be compared with published AA-LCR results, which grade written answers.
The architecture note is short. Subquadratic Sparse Attention (SSA) is the company’s linear-cost attention. A decision head then maps the resulting representations onto the allowed answers. The blog names four agent-loop roles: routing, tool picking, checking, and stopping.
The product is Public Preview, first come first served. The homepage and the post offer 100 million free input tokens to the first 10,000 sign-ups. After capacity, the blog says, new users may join a waitlist. The post’s speed line is 80 ms with almost no context, 250 ms at 32k, 660 ms at 100k, and 2.5 seconds at 384k. It also says latency will vary while the platform is tuned.
On October 12, 2026, at 19:26 UTC, Whedon posted that the service had processed 11.4 billion tokens in the previous hour and nearly 7 billion in the previous half hour, inside the first three hours. A follow-up at 19:48 UTC that day said configurations had changed for the request-size mix. Those two posts are traffic notes.
Opened October 13, 2026 UTC, the Decision Index space headers v0.3.2, 149 models, Jev jev-1.13.0, updated 2026-10-12. The Full score chart prints Jev at 60.1. A search of that page for Snap returned no row. 61.7 stays the company’s 0.2.1 print.
Alexander Whedon posted SubQ Snap 1.1 on October 22, 2026, at 19:48 UTC. The text names a tie for number one on the Artificial Analysis Decision Index, +7.2 points against Jev across 16 enterprise benchmarks, 96.2 on When2Call up from 75.2, 81% fewer confident errors, $0.04 per million input tokens, output free, and 384k context. A follow-up repeats the 7.2 and 5.5 point gains against Jev 1.14 and Snap 1.0. At 19:56 UTC he wrote that the preview slug routes to subq-snap-1.1.
@ArtificialAnlys, 20:02 UTC the same day, printed Snap 1.1 at 65 on Artificial Analysis Decision Index v1.0, tied with d1-ultra and Elephant 9B, eight points ahead of Jev 1.14 at 57. Non-hallucination is 86%, second to Elephant 9B at 93%, ahead of d1-ultra at 85%. Accuracy is 20% against 32% and 12%. HLE-MultiChoice is 18%, level with Elephant 9B. Cost is $30 per 1M tasks against $32 and $8.4. Context is 384K against 32K and 16K. 65 is not the October 12 61.7 cell.
Docs, opened October 23, name subq-snap-1.1 as the default Decision model, Public preview, input $0.04 per million tokens, output free. The Decisions examples use that id. A request may hold 1 to 128 questions. The window is still 384K tokens (393,216). $0.04 is not the October 12 $0.05 line.
This desk, Date headers 02:23:14 GMT to 02:23:20 GMT October 23. GET https://api.subq.ai/v1/models with no key returned HTTP 200, Content-Length 236, two ids: subq-snap-1.1 and subq-snap-preview. A POST with no key returned HTTP 401. Three POSTs with a key returned HTTP 200. Urgency on subq-snap-preview returned noul 0.999, 124 input tokens, 0 output, desk clock 1,533 ms, response model subq-snap-preview. The three-question call chose technical at confidence 0.95, frustration 0.99 at confidence 0.98, urgency noul 1.0, 211 input tokens, desk clock 1,324 ms. Urgency on subq-snap-1.1 returned noul 0.999, 124 input tokens, desk clock 1,101 ms, response model subq-snap-1.1. The two urgency nouls matched to six decimal places. At $0.04 per million input tokens, 124 tokens is $0.00000496 and 211 tokens is $0.00000844.
Other hosted calls of this shape are on decision models like Jev.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is Alexander Whedon, @alex_whedon, October 12, 2026, 16:03 UTC, status 2109676361371824229. The bio on that post reads Co-Founder, CEO at @subquadratic. The text introduces SubQ Snap as a long-context decision model. It says compute per token stays constant as context grows. It prints 384K tokens of multimodal context, 12x larger than Jev; 61.7 on Decision Index v0.2.1 vs Jev's 57.9, outperforming on 13 of 16 enterprise benchmarks; 100K tokens and a decision in under one second; $0.05/MM input tokens, output free. Available today. First 10,000 sign-ups get 100M free input tokens. Capacity will be limited. Latency will vary. Model speeds: 80ms for minimal context, 250ms for 32k, 660ms for 100k, and 2.5 seconds for 384k. The visit link resolved to https://subq.ai/. The attachment is a video of 76,367 ms. We did not transcribe a score from it. The share image is that video's still.
@subquadratic quoted the post at 16:05 UTC on October 12, status 2109676867439739261.
Blog https://subq.ai/introducing-subq-snap, dated October 12, 2026. TL;DR: up to 384K tokens; 100K tokens in about one second; text and images; Decision RULER 96.3 at 1MM tokens; Decision LCR 74.0 against Jev 48.0% chunked; 16 Decision Index enterprise tasks average 67.6 versus 61.6, lead on 13 of 16; Decision Index public suite 61.7 versus 57.9, all requests under 34k tokens; four of five subareas, arts & human taste +12.11, language understanding +7.46, retrieval & classification +6.27, tools & automation +3.39; $0.05 per million input tokens, output free. The Decision Index paragraph also prints 61.70 and 57.90. The enterprise table's Average (16) row is 67.54 against 61.6, gain +5.96. The homepage chart prints 67.5 versus 61.6 (+6.0). Public Preview, first come first served. A waitlist after capacity. SubQ 1.1 Small is Early Access. [email protected].
Decision RULER table on that blog: 32K Snap 99.8 / Jev 94.1; 64K 99.8 / N/A; 128K 99.5 / N/A; 256K 99.8 / N/A. The body says Snap scores between 96.3 and 99.8 from 32K to 1M tokens. The teaser image alt lists 99.8, 99.8, 99.5, 99.8, 99.3 and 96.3 from 32K to 1M. Decision LCR table: 72K-115K, Snap one-shot 74.0, Jev chunked 48.0. The body says 100 AA-LCR questions became 118 decisions, each run three times with different option orderings, credit only if every part is correct, chance 19.2, documents removed 24-25. It says Decision LCR scores cannot be compared with published AA-LCR results. Nimble Suite 13: Snap 82.9%, Jev 76.0%, 3,880 human-labelled examples from 13 public datasets.
Enterprise table, Snap / Jev / gain: CLINC150+OOS 91.5 / 89.2 / +2.3; BANKING77 80.6 / 79.5 / +1.1; PhishNChips 47.1 / 25.1 / +22.0; Amazon ESCI 48.7 / 43.8 / +4.9; BRIGHT 44.3 / 40.6 / +3.7; RAGTruth 64.8 / 51.3 / +13.5; HoVer 54.2 / 45.7 / +8.5; ContractNLI 70.7 / 59.1 / +11.6; FinEntity 92.8 / 80.8 / +12.0; NLI4CT 72.8 / 69.0 / +3.8; CRUXEval 73.8 / 57.1 / +16.7; ToolRet 63.1 / 59.9 / +3.2; BFCL 96.2 / 94.3 / +1.9; API-Bank 86.4 / 88.0 / -1.6; When2Call 67.0 / 74.6 / -7.6; ACOS 26.7 / 27.3 / -0.6.
Homepage https://subq.ai/, opened October 13, 2026 UTC. 384K context. Free 100M tokens for the first 10,000 users. Input $0.05 per 1M tokens, output free. Console link https://console.subq.ai.
Docs, opened the same day. Models page: subq-snap-preview, Public preview, Decisions endpoint, text in and out, image input only, input $0.05 / 1M tokens, output free. subq-preview is Early access for Chat Completions and Responses. Decisions API: POST /v1/systemone. Required fields model, state, questions. Context window 384K-token (393,216). Over the limit, or close to it, is 400. Images optional, up to 4, PNG JPEG or WebP, 4 MiB decoded each, 8 MiB decoded total, about 64-256 image tokens each, request body up to 13 MiB. Video is not supported. Noul, choice with 2-255 options, score with 2-10 levels. Choice and score return confidence. usage.output_tokens is always 0. Errors: 400, 401, 402, 403, 422, 429, 5xx. Quickstart signs up at console.subq.ai, env SUBQ_API_KEY, and says a request without credits returns 402. API overview base URL https://api.subq.ai/v1. Models API example lists both subq-preview and subq-snap-preview.
GET https://api.subq.ai/v1/models, no key, October 13, 2026 UTC. HTTP 200. One id, subq-snap-preview, owned_by subquadratic. An empty POST to /v1/systemone returned HTTP 422, error_type validation_error, message state: must have required property 'state'. A POST with model subq-snap-preview, a short state, and one noul, still with no key, returned HTTP 401, authentication_error, Missing API key. The same body with Authorization Bearer sk-test returned HTTP 401, Missing or invalid API key. We did not create a console account.
Later on October 13, 2026, at about 03:00 UTC, GET /v1/models with a bearer key kept in the environment returned HTTP 200. The list was the same one id, subq-snap-preview, name SubQ Snap Preview, owned_by subquadratic, created 0. It still did not list subq-preview. Content-Length was 135 with the key and without it.
Two POSTs to /v1/systemone with model subq-snap-preview returned HTTP 200. The state was the Stripe sentence this desk already used on October 6 with Jev and with OpenAI Decisions. This page does not print the request JSON. No image went out. The noul questions included true and false criteria, which the SubQ docs example includes. Date headers ran from 03:00:20 GMT to 03:00:23 GMT.
Urgency alone, instruction "Does this message express urgency?", returned noul 0.98. Usage was 119 input tokens and 0 output tokens. The response had no processing-time header. The clock on this desk, from send to the full body, was 1,195 ms.
The three-question call chose technical, confidence 0.90, with probabilities 0.93 on technical, 0.06 on billing, and 0.01 on sales. Frustration scored 0.98, confidence 0.93, with 0.95 of the weight on Frustrated but civil, 0.04 on Calm, just stating facts, and 0.01 on Very angry, strong language. The urgency noul was 0.98. Usage was 235 input tokens and 0 output tokens. The clock was 2,291 ms. Both responses named model subq-snap-preview. usage.output_tokens was 0. At $0.05 per million input tokens, 119 tokens is $0.00000595 and 235 tokens is $0.00001175. Neither body included an invoice line.
Hugging Face Decision Index space, opened October 13, 2026 UTC. Header: v0.3.2, 149 models, Jev jev-1.13.0, updated 2026-10-12. The Full score chart prints Jev at 60.1. A search of that page for Snap, SubQ, and 61.7 returned no Snap row.
Whedon, October 12, 2026, 19:26 UTC, status 2109727544132210730: 11.4B tokens in the last hour and nearly 7B in the last half hour, within the first three hours after launch. Latency spikes possible. A reply at 19:48 UTC that day, status 2109732974640173098, says configurations changed for the request size distribution. @subquadratic quoted the traffic post, status 2109730707593408792, "Builders are building."
Alexander Whedon, @alex_whedon, October 22, 2026, 19:48 UTC, status 2113356853308469444. The text names SubQ Snap 1.1, tied for number one on the Artificial Analysis Decision Index, +7.2 points against Jev across 16 enterprise benchmarks, 96.2 on When2Call up from 75.2, 81% fewer confident errors, $0.04 per 1M input tokens, output free, 384k context. A follow-up, 2113356859977425222, repeats the 7.2 and 5.5 point gains against Jev 1.14 and Snap 1.0, and When2Call 96.2 from 75.2. At 19:56 UTC, status 2113358899839684746, he wrote that the preview slug routes to subq-snap-1.1. @subquadratic quoted the launch at 20:05 UTC, status 2113361237560193363.
@ArtificialAnlys, 20:02 UTC October 22, status 2113360449156890662. SubQ Snap 1.1 scores 65 on Artificial Analysis Decision Index v1.0, tied with d1-ultra and Elephant 9B, eight points ahead of Jev 1.14 at 57. Non-hallucination 86%, second to Elephant 9B at 93%, ahead of d1-ultra at 85%. Accuracy 20% against 32% for d1-ultra and 12% for Elephant 9B. HLE-MultiChoice 18%, level with Elephant 9B, ahead of d1-ultra at 15%. Cost $30 per 1M tasks against $32 for d1-ultra and $8.4 for Elephant 9B. Context 384K against 32K and 16K.
Docs, opened October 23. Models page names subq-snap-1.1 as the default Decision model, Public preview, input $0.04 per million tokens, output free. Decisions API examples use model subq-snap-1.1. Questions per request 1 to 128. Window still 384K-token (393,216).
Date headers 02:23:14 GMT to 02:23:20 GMT October 23. GET https://api.subq.ai/v1/models with no key returned HTTP 200, Content-Length 236, two ids: subq-snap-1.1, name SubQ Snap 1.1, and subq-snap-preview, name SubQ Snap Preview, both owned_by subquadratic. GET with a bearer key returned the same list. A POST with no key returned HTTP 401, authentication_error, Missing API key. Three POSTs with a key returned HTTP 200. Urgency on subq-snap-preview, noul 0.999, 124 input tokens, 0 output, desk clock 1,533 ms, response model subq-snap-preview. Three questions on that id chose technical, confidence 0.95, probabilities technical 0.97, billing 0.03, sales 0.00. Frustration scored 0.99, confidence 0.98, 0.98 of the weight on Frustrated but civil. Urgency noul 1.0. Usage 211 input tokens, 0 output. Desk clock 1,324 ms. Response model subq-snap-preview. Urgency on subq-snap-1.1, noul 0.999, 124 input tokens, 0 output, desk clock 1,101 ms, response model subq-snap-1.1. The two urgency nouls matched to six decimal places. At $0.04 per million input tokens, 124 tokens is $0.00000496 and 211 tokens is $0.00000844.
Compare
61.7 and 57.9 are Subquadratic's Decision Index 0.2.1 public-suite cells. The September 28 board file still lists Jev at 57.91. Blink v0.3's self-run of that edition is 57.48 against 57.91. Fastino's GLiDE self-run is 64.81 against 57.91. Cloudflare's October 3 chart prints Clef at 61.2, captioned self-reported. Those four prints are separate from one another. The live space opened October 13, 2026 UTC is edition 0.3.2, Jev 60.1. d3's 64.7 is edition 0.3.1. None of those is Snap's 61.7.
The 16-task average 67.54, the TL;DR 67.6, and the homepage 67.5 are the same company table at three roundings. BANKING77 80.6 is that table. Clef-omni's BANKING77 macro-F1 94.8 is another post. RAGTruth 64.8 against 51.3 is this table. Cloudflare's RAGTruth F1 and Perplexity's 1,500-row accuracy are other tests. When2Call 67.0 against 74.6 is this table. Clef's When2Call 72.37 and Jev 80.97 are Cloudflare's.
Nimble 82.9% against 76.0% is Subquadratic's 13-set mean. Bespoke's published Jev cell of 76.0% is the same printed number Ollama later left on its 3,880-decision table, where Nimble itself is 75.7%. The 324-row Nimble holdout stays 90.12% against 93.21%.
Decision LCR 74.0 against 48.0 is the company's conversion of AA-LCR. The blog says it is not comparable with published AA-LCR scores. Decision RULER 96.3 at 1M tokens is that same blog.
$0.05 per million input tokens is the Snap rate card. TypeSafe's models page lists Jev at $0.042. TypeLLM's homepage lists $0.05 for input and $0.50 for thinking. OpenAI's Decisions guide lists $0.10. Those are separate cards. The 80ms / 250ms / 660ms / 2.5s line is the October 12 post. The October 13, 2026, 03:00 UTC desk clocks on the Stripe sentence are 1,195 ms for urgency alone and 2,291 ms for three questions, send to full body, no processing-time header. One call is not a ranking. The Snap noul questions included true and false criteria the October 6 Jev and Decisions calls did not. On that earlier pair, urgency was 0.98 on Jev and 0.99 on Decisions, and both chose technical, confidence 0.74 and 0.99, frustration 1.0. The October 7 and October 8 ten-call medians stay on the OpenAI Jev page.
65 is Artificial Analysis Decision Index v1.0, October 22 thread, tied with d1-ultra and Elephant 9B. 61.7 is the October 12 company Decision Index 0.2.1 cell. Do not merge those two. When2Call 96.2 is the October 22 post. When2Call 67.0 against 74.6 is the October 12 table. $0.04 is the October 22 docs rate. $0.05 is the October 12 card.
Date headers 02:23:17 GMT to 02:23:20 GMT October 23, noul 0.999 on both ids, three-question technical 0.95, desk clocks 1,533 ms, 1,324 ms, and 1,101 ms. Those three calls are not a ranking against the October 13 clocks. The preview slug still named subq-snap-preview in the response. The 1.1 id named subq-snap-1.1. The two urgency nouls matched.
Terms
- SubQ Snap
- Subquadratic's hosted decision model. Docs opened October 23 name subq-snap-1.1 at $0.04 per million input tokens, output free. GET /v1/models listed subq-snap-1.1 and subq-snap-preview. Date headers 02:23:17 GMT to 02:23:20 GMT October 23, three POSTs from this desk returned HTTP 200. The October 12 blog prints Decision Index 0.2.1 at 61.7 against a Jev cell of 57.9.
- subq-snap-1.1
- Pinned Decision model id on the Models page opened October 23. Whedon wrote that the preview slug routes to it. A POST with that id on October 23 named subq-snap-1.1 in the response. Artificial Analysis Decision Index prints 65, tied with d1-ultra and Elephant 9B.
- Stripe ticket
- The October 6 desk sentence Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP. On October 13, 2026, at about 03:00 UTC, Snap returned urgency noul 0.98, 119 input tokens, 0 output, desk clock 1,195 ms, then chose technical at confidence 0.90, score 0.98, noul 0.98, 235 input tokens, desk clock 2,291 ms.
- 384K tokens
- The Decisions docs window, 393,216 tokens. The October 12 post calls that 12 times Jev. TypeSafe's models page lists 32k for state plus the longest question and 64k for the whole request.
- Decision LCR
- Subquadratic's conversion of Artificial Analysis AA-LCR into 118 multiple-choice decisions over 72K-115K token document sets. The blog prints Snap 74.0 one-shot against Jev 48.0 chunked, and says the scores cannot be compared with published AA-LCR results.
- 100M free input tokens
- The homepage and the October 12 post offer 100 million free input tokens to the first 10,000 sign-ups. The blog calls the product Public Preview, first come first served, with a waitlist after capacity.
Sources
- Alexander Whedon, October 12
- Subquadratic, October 12
- Alexander Whedon, launch traffic
- Alexander Whedon, latency note
- Subquadratic, builders note
- SubQ Snap homepage
- Introducing SubQ Snap
- SubQ docs, Quickstart
- SubQ docs, Models
- SubQ docs, Decisions API
- SubQ docs, API overview
- SubQ docs, Models API
- SubQ console
- Decision Index space
- Alexander Whedon, Snap 1.1
- Alexander Whedon, When2Call 96.2
- Alexander Whedon, preview slug
- Subquadratic, Snap 1.1
- Artificial Analysis, Snap 1.1
- SubQ docs, Models, opened October 23