AI model
Z.ai language model, the first natively multimodal entry in the GLM-5 series, with 320B total and 18B active parameters via hybrid sparse/linear attention.
NVIDIA and Princeton's Dao AI Lab cache SSM inputs instead of state, targeting a decode step where SSM layers are the largest component up to 100K tokens. ReplaySSM is already integrated into TensorRT-LLM, vLLM and FlashInfer, NVIDIA says.
Reality
Anthropic launched Claude Haiku 5.5 at an average price about 75% below Haiku 4.5. The saving varies widely with prompt length, so teams moving classification, support or query traffic need to price their own requests before they switch models.
Perspective Coverage
22 publishersReality
Z.ai says the base model did not change between GLM-5.2 and GLM-5.3, so the coding jump and the doubled exploitation score come out of the same post-training run. Security teams inherit the second half.
Perspective Coverage
4 publishersReality
Cantina released apex-flash-1, an open-weights vulnerability-research model it says solved 40 of 60 tasks for $2.38, against $74.68 for Claude Opus 5 High. There is no hosted endpoint, so teams download the 321-billion-parameter weights, pay for their own inference and verify the numbers themselves.
Reality
Cantina released Apex Flash-1, an open-weights security model fine-tuned on 50 of its own vulnerability cases. Its only performance evidence is a 60-task evaluation the company ran itself, and a separate build is modified to refuse fewer requests.
Reality
Anthropic says Zhipu's freely downloadable GLM-5.3 built working V8 exploits in 50 of 410 tries, against 56 for its own restricted Claude Mythos Preview. With the weights public, its safeguards come off cheaply, so a lab that restricts its own model no longer keeps the capability out of reach.
Perspective Coverage
4 publishersReality
FORGE's developer computed every screen of a multi-agent research app from each run's event log, so a simulated run looked identical to a real one. That let real agents replace the simulator with no UI changes, and it let a default simulated run answer the wrong question with confidence.
Reality
The Information puts the price at about 86 times run rate. Reuters says nothing has closed. Either way, teams that pull weights by repo ID now have a silicon vendor in the artifact path.
Perspective Coverage
12 publishersReality
GLM-5.3-Flash leads agentic terminal work, DeepSeek V4 Flash is billed as the cheapest per token, and a 2.52B MiniCPM5-2B runs locally under Apache 2.0. The comparison flags most of those numbers as vendor-reported.
Reality
Z.ai's 320-billion-parameter model activates 18 billion per token and ships under MIT, so a buyer can download it and measure for themselves. Every capability figure published so far comes from Z.ai's own launch materials.
Reality
One MIT-licensed proxy on localhost is enough to serve OpenAI's own desktop client from self-hosted models. The models that fail there fail on Codex's tool-call format. One of three tested got lost.
Reality
Z.ai's GLM 5.3 rejects every way of switching thinking off and defaults to max effort, where 33 of 33 checkable tasks came back right at $0.00468 per correct answer against $0.01173 from GLM 5.2.
Reality
The model string keeps working after the cutover, so a pinned request comes back from V4.1-Flash. By the figures in the dev.to write-up, that model scores 90.6 on Terminal-Bench 2.1 and 42.3 on SimpleQA, against V4-Pro's 55.2.
Reality
Artificial Analysis scores GLM-5.3-Flash 42 at $0.25 a task and Kimi K3 44 at $2.00 a task. At eight-to-one on price, a two-point composite gap settles nothing, and the cost of an hour of human review decides it.
Reality
Investors bid the last of China's four little dragons of AI chipmaking up 206 per cent on its Shanghai debut, chasing the nearly 60 per cent of the domestic AI accelerator market that foreign suppliers held in 2025.
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
The MIT-licensed 320B model card claims it beats GLM-5.2 at a tenth of the price and approaches Claude Opus 4.8 on coding, but it names no dollar rate, and the comparisons are largely the vendor's own.
Reality
Google held its token rates flat, but the measured cost of an Intelligence Index task rose from $0.40 to $0.58 while the score still trails Fable and Sol. Picking a model for agent work is now a pricing exercise.
Reality
The OpenRouter rate card puts nearly 16x between the two models. The finished-task bill came in at 3.4x on one coding spec and 30x on a logic puzzle, and that gap is why the task, not the token, is the unit worth pricing.
Reality
Zhipu's open-weight MoE ties Claude Opus 4.8 on one composite index at roughly a fortieth of its list rate. At that price gap, the practical question stops being which model is better. It becomes how many attempts your router can afford.
Reality