Du wirst angemeldet...

Bitte warte, während wir deine Anmeldung überprüfen

Community-Newsletter

AI developer tools · What shipped

For a senior engineer who already reads HN. Real changes in AI developer tools today: releases with version numbers, papers with benchmarks, repos that crossed a threshold worth knowing. Skip hype threads, pre-announcement leaks, and recycled summaries. Always link primary sources.

Von Marius BongartsTech99 Ausgaben
Ausgaben
1 / 99
Über Nacht von KI aus öffentlichen Quellen erstellt, täglich aktualisiert.
AI developer tools · What shipped
Sonntag, 25. Oktober 2026
AI developer tools · What shipped

Benchmark snapshot: agent reasoning, voice SDK, and frontier benchmarks

1 Min. Lesezeit

Mistral Voxtral 3

Voice agents just got a production-ready SDK.

Mistral shipped Voxtral 3—speech-to-text, text-to-speech, and a voice agent framework in one stack [Quelle: Mistral]. Voxtral Small Transcribe 3 hits 4.0% word error rate across 50+ languages, beating Whisper on the FLEURS benchmark. The TTS engine adds emotional synthesis for 35+ dialects at $13 per million characters—undercutting OpenAI. Deploy on Mistral's EU hosting or self-host; the SDK handles real-time transcription, multi-step tool calling, and voice cloning from 30 minutes of audio.

Speech reasoning just became as integrated as text.

Benchmark roundup

Four specialized benchmarks just dropped fresh results.

Artificial Analysis released GDP.pdf, testing professional document-production: Meta's Muse Wonder Full leads at 32.6%, with GPT-6 Astra and GPT-6.1 Sol both at 31.0% [Quelle: BenchLM]. HLE-Verified, measuring frontier reasoning on 1,811 expert questions, puts Gemini 3.8 Flash at 54.9% with GPT-5.6 Sol at 54.5% [Quelle: BenchLM]. Vending-Bench 2 simulates a year-long autonomous vending business: GPT-6 Astra finishes with $15,514.70, testing long-horizon agent coherence [Quelle: BenchLM]. Blueprint-Bench 2 measures floor-plan generation from photos—Gemini 4 Argon leads at 54.4% [Quelle: BenchLM].

Agent and spatial reasoning now have quantified leaderboards.

Haiku 5.5 cost-capability matrix

Anthropic's five Haiku tiers now tile the cost-versus-speed plane.

Following yesterday's update, new pricing details show Haiku 5.5 Low at $0.02 per task with 6.68s time-to-first-token, while Max reaches intelligence score 43 at 236 tokens/sec and $0.12/input pricing [Quelle: Artificial Analysis]. Pricing spreads 8.7x across variants, letting teams dial per-workload efficiency instead of forcing one tier across all inference. The Intelligence Index v4.3.2 now measures performance across 10 evaluations spanning finance, legal, healthcare, and engineering—finer resolution than prior snapshots.

Commodity inference just became workload-specific.

Quellen
GDP.pdf Leaderboard & Scores — October 2026 | BenchLM.ai
GDP.pdf Leaderboard & Scores — October 2026 | BenchLM.ai
21 hours ago ... Model updatesRelease timelineAI model deprecationsAI RaceLLM pricingPrice vs performanceLLM speedAgent benchmarksPricing trendsLLM statistics. Research ...
benchlm.ai
KI-Zusammenfassung

Artificial Analysis released GDP.pdf benchmark results on October 23, 2026, evaluating 22 AI models on professional document-production tasks as part of its Intelligence Index v4.3. Meta's Muse Wonder Full leads with 32.6%, followed by OpenAI's GPT-6 Astra and GPT-6.1 Sol both at 31.0%. The benchmark measures success on professional knowledge work deliverables and currently serves as display-only reference data, not yet factored into overall model rankings. Results span models from Meta, OpenAI, Anthropic, Google, and others, with scores ranging from 32.6% down to 11.2%.

Quelle öffnen
HLE-Verified Leaderboard & Scores — October 2026 - BenchLM.ai
HLE-Verified Leaderboard & Scores — October 2026 - BenchLM.ai
20 hours ago ... Model updatesRelease timelineAI model deprecationsAI RaceLLM pricingPrice vs performanceLLM speedAgent benchmarksPricing trendsLLM statistics. Research ...
benchlm.ai
KI-Zusammenfassung

Gemini 3.8 Flash leads the HLE-Verified benchmark with a score of 54.9%, followed by GPT-5.6 Sol at 54.5% and Claude Opus 5 at 54.4%. The HLE-Verified benchmark, updated October 23, 2026, contains 1,811 verified expert reasoning questions and removes uncertain items from the original Humanity's Last Exam set. Six models have published results on this frontier multidisciplinary expert reasoning benchmark, which is currently displayed as a reference metric in the BenchLM catalog but excluded from overall ranking calculations.

Quelle öffnen
Claude Haiku 5.5: Release Intelligence, Performance & Price
Claude Haiku 5.5: Release Intelligence, Performance & Price
15 hours ago ... The Claude Haiku 5.5 release offers 5 models, each with different intelligence, performance ... AI Agents · Evaluations. Products. Optima · MicroEvals · Model ...
artificialanalysis.ai
KI-Zusammenfassung

Artificial Analysis released Intelligence Index v4.3.2, a comprehensive benchmark incorporating 10 evaluations including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The index measures AI model performance across specific capabilities and industries including Finance & Accounting, Strategy & Operations, Legal, Healthcare & Medical, Engineering, and Economics. Claude Haiku 5.5 is highlighted as positioned in the most attractive quadrant for cost-efficiency per Intelligence Index task, with the benchmark also introducing AA-PersonalityBench for assessing model personality traits across 8 dimensions.

Quelle öffnen
Vending-Bench 2 Leaderboard & Scores — October 2026
Vending-Bench 2 Leaderboard & Scores — October 2026
19 hours ago ... Model updatesRelease timelineAI model deprecationsAI RaceLLM pricingPrice vs performanceLLM speedAgent benchmarksPricing trendsLLM statistics. Research ...
benchlm.ai
KI-Zusammenfassung

Vending-Bench 2, a benchmark released in 2026, tests long-horizon agent coherence by simulating a one-year autonomous vending-machine business operation. The October 23, 2026 snapshot evaluated 71 AI models, with GPT-6 Astra leading at $15,514.70 arithmetic mean final bank balance, followed by GPT-6 Sol ($14,427.85) and Gemini 4 Argon ($13,718.16). The benchmark measures agents' ability to manage product sourcing, supplier negotiation, inventory, pricing, and customer requests over a 365-day simulation, with each model run reporting standard error metrics. Results are displayed for reference only and excluded from weighted overall rankings.

Quelle öffnen
Blueprint-Bench 2 Leaderboard & Scores — October 2026
Blueprint-Bench 2 Leaderboard & Scores — October 2026
20 hours ago ... Model updatesRelease timelineAI model deprecationsAI RaceLLM pricingPrice vs performanceLLM speedAgent benchmarksPricing trendsLLM statistics. Research ...
benchlm.ai
KI-Zusammenfassung

Blueprint-Bench 2, an agentic spatial reasoning benchmark measuring model performance on floor plan generation from apartment photographs, was updated October 23, 2026 with results from 33 models. Gemini 4 Argon leads at 54.4% normalized connectivity score, followed by Claude Opus 5.5 (51.2%) and GPT-6 Astra (49.7%). The benchmark evaluates how well models process roughly 20 interior photographs per apartment and draw structured floor plans, with scores normalized to a 0–100 scale. BenchLM currently displays Blueprint-Bench 2 for reference only and excludes it from overall model rankings pending Google's full methodology publication.

Quelle öffnen
Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages
Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages
4 hours ago ... Your voice AI, on your infrastructure ... Macro-average WER over the intersection of 45 languages in the FLEURS benchmark supported by all seven models.
mistral.ai
KI-Zusammenfassung

Mistral released Voxtral 3, a family of state-of-the-art speech-to-text and text-to-speech models with an accompanying voice agent SDK. Voxtral Small Transcribe 3 achieves state-of-the-art multilingual transcription accuracy across 50+ languages, outperforming Whisper, Scribe v2, and Gemini 3.5 Transcribe on the FLEURS benchmark with 4.0% word error rate in forced mode and 4.5% inferred. Voxtral Mini TTS 3 provides emotionally adaptive speech synthesis for 35+ languages and 50+ dialects at $13 per 1M characters, undercutting OpenAI TTS-1 and comparable competitors. The voice agent SDK integrates real-time transcription, reasoning via general models like Mistral Small, and speech generation into a single streaming pipeline, supporting end-of-turn detection, context biasing, voice cloning from 30+ minutes of audio, and complex multi-step tool calling, available via EU-hosted API or self-deployment.

Quelle öffnen
Über Nacht zusammengestellt von MorningMail.aiZugestellt um 06:10

Mehr aus Tech

Alle Newsletter aus Tech ansehen →