Topic
Performance comparisons published by interested parties, where harness, tooling, prompting and concurrency determine applicability.
AWS says Trane's engineers cut a 20-minute dashboard workflow to a 20-second question in three to four weeks. The multiple rests on timings Trane ran with its own technicians, and the durable part is where the tool calls execute.
Reality
PrismML's ternary build of Qwen3.8 27B keeps 98.2 percent of the full-precision benchmark average on both of the company's inconsistent scorecards, and the loss it does take is concentrated in knowledge and reasoning.
Reality
Alibaba's new omni model takes text, images, audio and video across a million-token context and answers in text plus function calls. Rendering and token budgeting stay in the stack you already run.
Reality
Google's September 15 release pairs a fast speech-to-speech model with one that narrates its own reasoning aloud while tools run. Clients now have to read interactionStatus to know when a turn is actually over.
Reality
DeAlignAI's downloadable FP8 build is the license working exactly as written, while its self-reported 320-of-320 HarmBench run remains unchecked by any outside researcher and measures compliance, not capability.
Reality
Z.ai says the model runs at a tenth the cost of its last one. The comparison an operator needs is against the API invoice they already pay, and the release does not make it.
Reality
Jalapeno, co-designed with Broadcom, starts landing in OpenAI's own data centers by the end of 2026. The only efficiency figure is OpenAI's, and the part is not for sale.
Reality
Inherent says its Faraday agent reproduced published findings better than Claude Opus 4.8 and GPT-5.5. The interesting number is not the parameter count but what the account leaves out.
Reality
The update adds a path selector and a two-tap convolution rather than layers, recovering most of the accuracy that tripling the drafter bought at 15.2% latency, by the vendor's own numbers.
Reality
Z.ai says GLM-5.3 edges Anthropic's restricted Mythos 5 at vulnerability discovery while losing badly at exploitation. On vendor numbers, the defensive half is commoditising first.
Reality
The Ultrafast preview runs GPT-5.6 Sol on Cerebras hardware for a hand-picked customer list. That makes capacity allocation, not model choice, the constraint your architecture has to survive.
Perspective Coverage
3 publishersReality