News and analysis

AI and infrastructure announcements, read against their primary sources, with what is confirmed kept apart from what is only claimed. Newest first. For prices, versions and dates in one place, see the reference trackers.

Latest news

DevGranite Speech 5.0 2B claims #2 open, the live board says 10thA read of IBM's Granite Speech 5.0 post against the model card and the live Open ASR Leaderboard: what holds, what shifts,…23 October 2026DevOpenAI’s 12-hour prompt cache pays off from the third requestGPT-5.6 and later can keep a cached prefix for 12 hours at a 2.5x write price. Worked break-even numbers, and the traps…23 October 2026DevAnthropic’s new API version rejects temperature, even at its defaultAnthropic's 2026-10-21 API version rejects sampling parameters and some citation and document placements, and Python 2.0.0 and TypeScript 1.0.0 send it by…22 October 2026DevTorchTPU 0.2.0 is out, and 96.5% of its runnable tests passTorchTPU 0.2.0 lets Transformers and Diffusers target a TPU with device_map or .to. The 96.5% pass rate counts only tests that run…20 October 2026DevSorting your prompts speeds up MoE batch jobs by 5% to 17%Sorting prompts by domain before offline MoE inference gained up to 17.3% on Qwen3-30B-A3B, but only about 5% with unlabeled text, on…20 October 2026DevNemotron 3.5 Super VL fits one DGX Spark, with fine printWe read both Super VL model cards and the NVFP4 file list. One Spark works on paper. The card's own BF16 versus…19 October 2026DevClaude’s 17x quantum gate cut is measured against the slowest circuitAnthropic says Claude Mythos 5 found an elliptic-curve circuit with 17x fewer gates for 12% more qubits. We read Table S3 of…19 October 2026DevDeepSeek’s search API caps every account at 5 requests a secondDeepSeek opened Search and Fetch on 17 October at $0.75 per 1,000 successful calls. We read the guide: one account gets 5…18 October 2026SysadminPostgreSQL 14 goes unsupported a week before PostgreSQL 19 shipsPostgreSQL 19 RC1 is out, GA slid to 19 November, and PostgreSQL 14's final release lands a week earlier. We read the…18 October 2026DevOpenAI’s free prompt voices only speak inside GPT-LiveDescribe a voice, get a voice id, pay nothing. We read OpenAI's custom voices guide and found where prompt voices can't go:…18 October 2026DevFLUX 3 Image Klein runs in 8GB, but not with the upsampler it wantsBlack Forest Labs says Klein runs on 8GB RTX cards. The INT8 generator and text encoder files already total 10.69 GB, and…17 October 2026DevDeepSeek swapped deepseek-flash today, and there’s no id to pin backOn 17 October DeepSeek put V4.1-Flash-1017 behind deepseek-flash, same price, bigger agent scores. The API offers no dated id for the old…17 October 2026DevGranite Switch 4.1 won’t install next to this month’s vLLMIBM released Granite Switch 4.1 on 15 October. The model cards pin vLLM below 0.28, the plugin allows up to 0.30, and…17 October 2026DevBeam’s weights are out, and vLLM support is still an open pull requestReflection published Beam 1.0 weights on 16 October: 1,002.68 GB in BF16, 293.65 GB in NVFP4, plain Apache 2.0. Transformers 5.20.1 loads…16 October 2026DevIs Opus 5.5 fast mode worth $40 per million output tokens?Since 15 October, fast mode on Claude Opus 5.5 needs no waitlist or beta header. We stacked every multiplier on the $8/$40…16 October 2026DevFine-tune on Muse Spark Open outputs and you owe Meta a credit lineMuse Spark Open went up on Hugging Face on 15 October under an MIT-style licence with three strings: a Built with credit…16 October 2026DevNemotron 3.5 Bolt says 24GB on the label and 28GB in the recipeNVIDIA released Nemotron 3.5 Bolt 74B-A9B on 15 October. The files the recipe loads total 24.75 GB, there is no 24 GB…15 October 2026SysadminClaude API traces won’t reach a collector on 4318 or a private IPAnthropic shipped OpenTelemetry trace export for the Claude API on 14 October. We read the delivery rules: port 443 only, public hostnames…15 October 2026DevMeta’s API sells Muse Wonder, not the Wonder Full that tops its chartMeta launched Muse Wonder and Muse Wonder Full on 12 October. We read the model card and the price page: the API…15 October 2026DevLFM2.5-24B-A2B beats Qwen3.6 at tools and loses SWE-Bench by 15 pointsLiquid AI released LFM2.5-24B-A2B with open weights. We read the full benchmark table and the license: strong at tool calls, weak at…14 October 2026DevRequests to gemini-3.7-flash now get 3.8 Flash, with no shutdown dateGoogle deprecated Gemini 3.7 Flash on 8 October and routes its traffic to 3.8 Flash, a month after calling 3.7 the efficiency…12 October 2026DevOpenAI’s Decisions API charges $0.10 in and nothing out, with no cacheDecisions returns a probability, a choice or a score instead of text, on gpt-6-luna only. It's $0.10 in and free out, but…11 October 2026DevGPT-6.1 Sol Ultrafast bills six times Sol and 1.2 times AstraUltrafast on GPT-6.1 Sol costs $12 and $60 per million tokens: six times standard Sol, three times Fast, and 1.2 times Astra…10 October 2026DevClaude Haiku 5.5 is $0.10 per million until a prompt passes 100KAnthropic released Claude Haiku 5.5 on 7 October at $0.10 and $0.50 per million tokens. The footnote matters more than the headline:…8 October 2026DevReflection Beam is a 501B open model with no weights yetReflection AI's Beam is a 501B mixture of experts with an Apache 2.0 promise, a waitlist and benchmark scores nobody outside can…6 October 2026DevKolibri-1 serves 1M tokens but was trained to 256KAleph Alpha released Kolibri-1 on 3 October 2026 under Apache 2.0. The headline says 1M tokens of context. The model card says…4 October 2026DevGemini 4 Argon costs $2 and $10 now, $4 and $20 laterGoogle listed Gemini 4 Argon at $2 input and $10 output per million tokens on 30 September 2026, with a later rate…2 October 2026DevGPT-6.1 Sol costs what GPT-6 Sol did, and only the cache got cheaperA fifth of Astra’s price, yes. Against GPT-6 Sol, the only price that moved is cached input, cut from $0.20 to $0.10.30 September 2026DevClaude Sonnet 5.5 keeps $2 and $10 and retires thinking disabledSame $2 and $10 price as Sonnet 5, benchmarks within two or three points of Opus 5.5, and a disabled thinking setting…29 September 2026DevMiMo-V2.6-Pro was quietly retrained to stop looping on tool callsSame model names, new weights. Xiaomi retrained MiMo-V2.6-Pro and Flash to stop repeating tool calls, shipped it to the API first and…28 September 2026

Older news, in the archive →

Explainers

Longer pieces on the networking and security topics people keep asking about.

DevClaude Fable 5 vs Grok 4.5: is 8x cheaper worth it?Two coding models, one released a week apart, sitting at opposite ends of the market. Claude Fable 5 is the benchmark king:…8 July 2026DevxAI Grok 4.5 is here: cheap, fast, coding-firstxAI shipped Grok 4.5 on July 8, and the number that matters is not a benchmark, it's the price: $2 in, $6…8 July 2026DevClaude Fable 5 is back: what each effort level costsClaude Fable 5 lasted three days. Launched June 9, suspended June 12 under a US export control order, and as of July…2 July 2026DevClaude Sonnet 5 vs Fable 5: is the top model worth 5x?Anthropic pulled off a strange double feature. On June 30 it launched Sonnet 5 at $2 in and $10 out per million…2 July 2026DevClaude Sonnet 5 vs Opus 4.8: cheaper, nearly as goodOn June 30 Anthropic shipped Claude Sonnet 5, made it the default in Claude Code, and priced it to make you look…1 July 2026SecurityProton Lumo 2.0 review: how private is it, really?Proton shipped Lumo 2.0 on June 30, and the headline writes itself: a private ChatGPT alternative, built in Europe, that got a…1 July 2026SecurityNo, Claude Mythos didn’t hack the NSA: what really happenedFor a few days in June 2026 the internet was sure an AI had hacked the NSA. Claude Mythos, Anthropic's frontier model,…22 June 2026DevGPT-5.6 vs GPT-5.5: what actually changedOpenAI started previewing GPT-5.6 in June 2026, and the first surprise is that GPT-5.6 isn't a model, it's three. The new naming…22 June 2026DevQwen 3.7 Max vs GLM-5.2: benchmarks, price and the catchAlibaba dropped Qwen 3.7 Max in May 2026, and within days the coding crowd was lining it up against GLM-5.2, the open-weight…22 June 2026Email & DNSHow to read a DMARC report: the aggregate XML, decodedSo you switched DMARC on, and now google.com mails you a little .xml.gz file every morning. You open one, meet a wall…20 June 2026SysadminLinux file permissions explained: rwx and 755You chmod a file, it works, you move on, and permissions stay a thing you never quite have to learn. Then a…20 June 2026SysadminMB vs MiB: why your 1 TB drive shows up as 931 GBYou buy a 1 TB drive, plug it in, and Windows cheerfully reports 931 GB. Nothing's broken and nobody robbed you, but…20 June 2026