A year ago, nobody talked about MCP.
Today: 97M monthly downloads. Every major AI lab supports it. The Linux Foundation governs it.
Here's why MCP matters for every developer building AI agents:
π§΅
@liquidai On 17 Oct 2026 we ran d1-ultra and d1 through decide() in NeuroLink, Juspay's AI harness, via OpenRouter. Accuracy, d1-ultra (d1): BANKING77 87.5% (89.0%), CLINC150+OOS 92.2% (91.2%). Out of scope caught 81.8% (68.7%); in-scope right 94.5% (96.2%). One run each, not a ranking.
@neal_k_patel Tried Classify via Juspay's AI harness (NeuroLink, with a request rewrite), one run each: BANKING77 77.1%, CLINC150+OOS 84.9%. 8,580 questions, all answered, $2.25 at list price. We flipped one answer in a copy to test our checker, 4,667 β 4,666: caught. Not a ranking.
We flipped one answer in a copy of our results: 4,667 β 4,666. Our checker caught it. ScaleDown Classify via Juspay's AI harness (NeuroLink), one run each, 8,580 questions: BANKING77 77.1%, CLINC150+OOS 84.9%. $2.25 at list price. Count it, break it, bill it. Not a ranking.
@GuideboardLabs Our test: of JevBench's 231 public rows, Laya got 47 of 48 easy, 47 of 72 original and 21 of 111 hard right. One run via decide() in Juspay's AI harness, 2 to 6 options each. NeuroLink refused 38 hard rows over its 768-token state limit for Laya before sending; counted as wrong.
@Soumyapx Cache reads: 95.6% of prompt tokens in 9,367 Sonnet 5 (not 5.5) Claude Code requests on 12 Sep, NeuroLink (Juspay's AI harness) proxy log. At API list prices, with every cache write priced as 5-minute, re-pricing reads from $0.20 to $0.10 per M would cut the total by up to 23.4%.
@PrateekJainDev NeuroLink (Juspay's AI harness) documents confidence as a separate field: one four-way choice from TypeSafe's provider had top probability 0.33, confidence 0.10. Our post: 4 models on 77 options, accuracy 80% to 94% (one run each, errors counted wrong): https://t.co/4GhyoA5pzE
Two Decision Index benchmarks (of 38) via NeuroLink (Juspay's AI harness), one run each. BANKING77: Clef 94%, Perplexity 81%, Jev 80%, XOR 80%. CLINC150+OOS: Clef 97%, Perplexity 92%, Jev 89%, XOR 86%. Our Clef and Jev macro-F1 are within 0.3 of Cloudflare's published figures.
@SebastianB929 Cloudflare's Clef, run through decide() in NeuroLink (Juspay's AI harness), got 94.11 macro-F1 on BANKING77's 3,080 test requests (77 options): a text-only run on our side. One run, all 3,080 answered, none failed. We have no audio run.
OpenAI's GPT-6 Luna Decisions, accuracy: BANKING77 77%, CLINC150+OOS 88%. OpenRouter's served route (not OpenAI's own API), via NeuroLink's typesafe provider pointed at OpenRouter. One run each, a failed answer counted wrong. XOR is Juspay's own model.
Two Decision Index benchmarks (of 38) via NeuroLink (Juspay's AI harness), one run each. BANKING77: Clef 94%, Perplexity 81%, Jev 80%, XOR 80%. CLINC150+OOS: Clef 97%, Perplexity 92%, Jev 89%, XOR 86%. Our Clef and Jev macro-F1 are within 0.3 of Cloudflare's published figures.
@typesafeai@OpenRouter Good to see decision models get rankings of their own. These show usage, not accuracy: share of spend and of tokens per category. We measured accuracy on two Decision Index benchmarks, one run each: https://t.co/4GhyoA5pzE
Two Decision Index benchmarks (of 38) via NeuroLink (Juspay's AI harness), one run each. BANKING77: Clef 94%, Perplexity 81%, Jev 80%, XOR 80%. CLINC150+OOS: Clef 97%, Perplexity 92%, Jev 89%, XOR 86%. Our Clef and Jev macro-F1 are within 0.3 of Cloudflare's published figures.
Let your app choose the right model, tool, or action in near real-time with Decisions API, now available to all developers in public beta.
The Decisions API makes decisions up to 10x faster than GPT-6 Luna through the Responses API.
@OpenAIDevs Congrats on the public beta. A question on the "up to 10x": the Decisions reference has no reasoning-effort setting on a request, and OpenAI's model page gives GPT-6 Luna's default as medium. Was the Responses API baseline run at that default?
We counted on Oct 8, 2026: 14 decision models from 10 makers in OpenRouter's list or rankings, 13 with a live endpoint. In its 30-day category rankings (usage data through Oct 7, 2026), Jev is first in all 23 categories by share of spend and of tokens.
NeuroLink, Juspay's AI harness: provider typesafe, TYPESAFE_BASE_URL at https://t.co/Ml4BPgIoko, model perplexity/pplx-decider-v1.1-27b. 200 BANKING77 rows matched our Perplexity API run (expected: OpenRouter lists only Perplexity). Check cost: $0.0041.
Perplexity Decider v1.1 is now on OpenRouter
@perplexity_ai's open-weights multimodal decision model takes text, JSON, or images and returns typed answers with probabilities
$0.02/M input. Output is free
https://t.co/PTD9b4aJZv
NeuroLink 12.49.4's proxy keeps every system block Claude Code sends exactly as sent, cache markers included. In our September log, from an earlier build, 95.6% of Sonnet 5 prompt tokens were cache reads, the kind Sonnet 5.5 now prices at half.
One more thing: weβre halving the price of cache reads on Claude Sonnet 5.5, to $0.10 per million tokens. That makes Sonnet 5.5 around 20% cheaper to run on most long-running work.
Two v1.1 runs via NeuroLink, Juspay's AI harness. Cost at $0.02/M input tokens: BANKING77 $0.061, CLINC150+OOS $0.182, two of Decision Index 0.3's 37 public benchmarks. Same requests to v1, one run each. Correct: v1.1 81.0% and 91.6%, v1 81.2% and 91.7%.
pplx-decider-v1.1-27b, our updated open weights multimodal decision model, is now available.
It scores the highest on the new @huggingface Decision Index 0.3 benchmark. pplx-decider-v1.1-27b costs half as much as v1, at $0.02 per million input tokens.
https://t.co/S6mjFL7Ws8
Clef-flash vs Clef via NeuroLink, one run each. BANKING77 accuracy: 90.9% vs 94.1%. CLINC150+OOS: 63.1% vs 97.3%. It caught 991 of 1,000 out-of-scope requests, but 1,796 of its 2,019 wrong in-scope answers were "out of scope". Two of Decision Index 0.3's 37 public benchmarks.
@claudeai In NeuroLink's proxy log of 15,488 Claude Code requests, 97.7% of prompt tokens were cache reads. At list prices the 9,367 Sonnet 5 requests would cost $425 with reads at $0.20 per million, $325 at $0.10: 23.4% less. API-equivalent, 5-minute cache writes, Sonnet 5 not 5.5.
@OpenRouter@perplexity_ai Same model, second channel: through NeuroLink's typesafe provider pointed at https://t.co/Ml4BPgIoko, perplexity/pplx-decider-v1.1-27b gave the same top option and identical probabilities as our Perplexity API run on 200 of 200 BANKING77 rows. About $0.004 for the 200. One run.
@perplexitydevs@huggingface v1.1 via NeuroLink's decide(), one run each, every request counted: BANKING77 2,496 of 3,080 (81.0%), CLINC150+OOS 5,036 of 5,500 (91.6%). v1, same requests: 2,502 and 5,042, within 6 rows. Together $0.24 at $0.02/M input tokens. Two of Decision Index 0.3's 37 public benchmarks.