AI News: Ultrafast Inference, Claude Watermarks and New Models

This week brought an unusually large amount of AI news: an ultrafast ChatGPT mode, Claude text watermarking and several new models. I will go through the announcements and what they might mean for users.

OpenAI is previewing an ultrafast GPT-5.6 Sol mode using Cerebras hardware. In a displayed dashboard-building demonstration, the ultrafast run finishes in 1 minute 50 seconds while the comparison run takes 12 minutes 20 seconds. This is one example, not a controlled measure of every task; the faster mode costs extra.

I usually launch many agents at once because each can take a long time, but keeping track of them adds context-switching costs. After briefly trying the preview, I think faster inference could let me work with fewer concurrent agents. Tool calls and local computing may then become a larger share of the wait.

Anthropic announced watermarking for future Claude outputs. I initially wonder how it works, whether it changes the text and what the implications are for tracking. I also question the EU transparency requirements cited in connection with the announcement.

An editor adds an update after Anthropic published more detail. The company says watermarking uses a keyed pattern in low-stakes word choices, rather than visible markers: the model still selects among possible words, but uses a key and preceding words to influence the random choice. The editor describes the stated goal of making the pattern detectable with the key.

The editor explains that lightly edited human writing offers fewer model-selected words in which to encode a watermark. Anthropic says code may contain watermarks in flexible language such as comments while claiming negligible effects on the actual code. The editor questions how much changing the selection process could affect output, and says the long-term effects remain to be seen.

I turn to Cursor’s move into SpaceX and the Grok 4.6 release. I view the model as much improved and close to leading systems on the comparisons shown. The video reports a price of $2 per million input tokens and $6 per million output tokens at the time.

Grok Bot offers a simplified agent interface: users need not inspect code or select a model and reasoning effort for each task. After about a week of use, I plan to move some workflows to it. I compare it to a more accessible version of OpenClaw.

I highlight individual threads as agents, integrations such as Slack, Google Docs and email, and the ability for agents to delegate and communicate. Their conversations are retained, so multiple agents can work toward a shared goal.

GLM-5.3 is the first of several open-model releases I discuss. The displayed comparison shows a substantial improvement over GLM-5.2 on Terminal Bench 3.0 and a 66.9 score on DeepSWE. I consider it competitive but below some leading models in the tests shown; these are reported benchmarks rather than a universal ranking.

DeepSeek-V4-Pro is another release. I cite a Terminal Bench result of 87.9 in the displayed comparison, close to Kimi K3 and Fable 5 at 88 in that table. Different benchmark versions and test conditions matter when comparing results.

The displayed DeepSeek API pricing separates peak and off-peak hours and cache hits from misses. For V4 Pro, the listed off-peak prices are $0.66 per million input tokens on a cache miss, $1.98 per million output tokens and $0.022 per million cached input tokens. The screen says those prices take effect August 16, 2026.

Meta’s Muse Glimmer is a 30-billion-parameter open-weight model intended for local agent workflows. Its displayed announcement lists an Apache 2.0 license. I note that it is not intended to compete with the largest frontier models and suggest some desktop GPUs might run it, subject to memory and configuration.

OpenAI’s Computer History feature can use selected computer activity to make future interactions more contextual and suggest automation. I say the feature is opt-in and allows controls over what to share. People may reasonably be cautious about recording computer activity.

The idea reminds me of Microsoft Recall, which drew strong privacy and security objections. I was interested in that feature and say I may try sharing a few apps with Computer History. My favorite release of the week is Grok Bot.