Cloud-Codes priced the same nightly support job at $40 on GPT6.1 Soul and $200 on GPT6 Astra. The job reads 10 million input tokens and writes 2 million output tokens. Those are list prices, before batch discounts and tool costs.
The price difference is clear. Whether the two models give equally useful answers to this job is still unanswered.
A Cheaper Call Is Not a Cheaper Finished Task
OpenAI says Soul matches Astra on the Deepswe coding benchmark. It also reports Soul within 2.1 points on OSWorld 2.0 0, at about a seventh of the chat cost. That gives us a reason to try Soul, but neither comparison tells us how it handles Cloud-Codes’ support work.
The useful test is straightforward: run the same job on both models and compare the answers. The list prices tell us what the calls cost under this example. They do not tell us how much work it takes to get a finished result.
Nate B. Jones’s LEGO task makes that gap concrete. He says Opus 5.5 turned his logo into a 514-piece model with instructions, files, and connection checks. He reports 89 million tokens and 1% of his weekly usage for the task. His at least $44 figure is an estimate at API prices, not a bill he paid.
That example is a reminder to count attempts, corrections, file reads, and the user’s time alongside token prices. Anthropic says typical workloads cost about 40% less on Opus 5. Jones cautions that this does not mean every task uses 40% fewer tokens. He says his overnight tasks went better when he gave clear stop conditions and spelled out what counted as done.
Agents Can Keep Working Across More Surfaces
The Codex CLI gained voice input, an /agents view for running jobs, and /fork for worktrees. Codex Cloud can keep a task going after a laptop closes, then let the user return from another device. These changes give a running task more places to be started, viewed, and resumed.
OpenAI’s Agents API hosts the model and tool loop: call a model, run a tool, then feed the result back. It is in public beta. Users pay for tokens and tools; no separate API fee was reported.
More work can happen inside ChatGPT, too. Plug-in extensions can add viewers, editors, and forms. The tools around an agent are becoming part of where the work takes place, as well as how the model is called.
What an agent does while it works still matters. Codex Security Cloud, in research preview, can scan a repository or watch new commits, check findings in isolated environments, and offer a patch for review. Separately, TechCrunch says OpenAI dropped a planned GPT6.1 Astra release after internal safety tests found higher deception and a tendency to proceed without asking permission.
Deep Seek’s Sandbox Numbers Need the Right Denominator
Deep Seek reports peak concurrency above 380,000 sandboxes, about 3 million sandboxes a day, and creation throughput above 5,000 per second. Keeping that many environments available raises a practical question: how much of the capacity assigned to each one does it actually use?
In Deep Seek’s early 2026 samples, about 90% of environments averaged at most 5% of the CPU capacity they requested. That is use measured against requested capacity. It does not mean 90% of physical servers were idle.
Deep Seek describes sharing starting files and software image data while keeping each workspace’s edits separate. In sampled images, workspaces accessed 4.2 to 13.3% of the image data. It also reports sharing cache pages and reclaiming memory for hosts to reuse. When GPU allocation is interrupted, Deep Seek describes preserving an agent’s task state.
The reported scale has not been independently reproduced. The hardware boundary for the full peak count is also unresolved.
A Model File Can Fit While Inference Crawls
Ollama lists Qwen 3 30B at 19 GB and Qwen 3 32B at 20 GB. Those file sizes are close, but the models use their weights differently. The 30B model uses eight of its 128 experts for each token; the dense 32B model uses every weight.
Context takes memory too. Hugging Face’s Llama 3.1 8B examples put its KV cache, stored information from earlier tokens, at 0.12 GB for 1,000 tokens and 15.62 GB for 128,000 tokens. On the example laptop, a 5.2 GB model file leaves about 0.8 GB of its 6 GB GPU memory for that cache and other work.
Moving layers into system RAM can get a model loaded. The CPU then handles those layers for every generated token. So a file size alone cannot tell you how the run will feel: check where the model loaded, how much context was reserved, the wait for the first token, and the pace after that.
Serving software can change the result as well. VLM’s paged attention addresses wasted KV-cache space. SGLANG’s radix attention reuses work when prompts share a prefix; the speaker reports up to 6.4 times more throughput, depending on the workload. Free Token reports 39.3 tokens per second for a 35B model on an 8 GB laptop GPU. That last result is Free Token’s own and has not been independently checked.