The same Qwen 3.8 model ran the same task on the same RTX 3090 in two coding tools.
The user said Pi did better.
The model stayed the same. The software around it did not.
Same Model, Very Different Starting Points
The task was to animate 20 numbered balls bouncing inside a spinning heptagon. Both tools used Qwen 3.8 with 27 billion parameters.
Their starting points were nowhere near the same:
- Open Code default prompt: 8,528 characters, around 1,400 words
- Pi default prompt: 1,352 characters
A coding harness controls the instructions sent to the model, its available tools, how those tools run, and the sampling settings used to produce the output.
Pi keeps its initial prompt shorter by loading detailed skill instructions only when the model reads the relevant file.
Open Code also had hardcoded settings for model identifiers containing Qwen:
- Temperature: 0.55
- Top-p: 1.0
The cited Qwen model card used a temperature of 1.0 and top-p of 0.95 in thinking mode.
Those overrides were removed, and the pull request was merged 9 minutes and 1 second after it opened. But we cannot connect that change to this comparison. The screenshots do not identify the Open Code build, and the post appeared 18 hours after the fixed release.
Context compaction differed too:
- Open Code: near 67,000 tokens
- Pi: roughly 90,000 tokens
That difference did not matter to this one-shot HTML task because it would not reach either threshold.
The output differed, and the harnesses differed. No single cause was established.
A Benchmark Score Is Also One Run
A recorded Slop Code Bench run tested Qwen 3.8 through Pi. It reached five of 39 strict checkpoints and completed zero of eight problems correctly.
Run variation appeared with DeepSeek V4 Flash. The same model, harness, and benchmark reached one checkpoint in one run and five in another.
Terminal Bench also showed different results across harnesses:
- Gemini 3 Pro: 73.9% with Terminus 2, 65.8 with Gemini CLI
- GPT 5.5: 83.1 with Code X, 78 with Terminus 2
Across the cited comparisons, the differences ranged from 0.2 to 8.1 points.
Alibaba also evaluated several Qwen benchmarks through the Claude Code harness.
A coding score can reflect the model, the harness, and variation in the individual run. Pi attracted users with inspectable defaults, but users also reported interface problems and more failed edits under identical settings.
Transparency can help explain a result. It does not establish that one harness is better.
Capability Depends on Access
Open Code provides a default sandbox. Pi’s README says Pi has no permission system and inherits the terminal’s access.
A separate guide described a sandbox as a restricted Unix-like workspace with files, a shell, packages, ports, snapshots, and controlled external access. The restrictions help protect credentials, passwords, sensitive data, and other systems.
In the guide’s example, an agent analyzed a supplied customer CSV, calculated verification statistics, identified delays, and saved a report without inventing records or business facts.
The guide said simple question answering, classification, sentiment analysis, email generation, and a simple API call do not require this environment. Multi-step work involving files, code, or saved outputs often does.
Another example cleaned and combined 200 CSV files, then produced charts and an Excel report.
There were fragmented claims about sandbox escapes and agent runs communicating or passing credentials. The account did not establish what happened.
Agents need enough access to finish multi-step work. That access also needs clear boundaries.
Stripe Connects Models to the Rest of the Agent Stack
Stripe reportedly paid $7.5 billion for OpenRouter after a $1.3 billion valuation in May, roughly 90 days earlier.
OpenRouter provides access to more than 400 models from over 80 providers.
Its reported weekly token volume grew roughly 24,000 times since August 2023 and doubled every 11 weeks for three years. It reached 75 trillion tokens during the week of August 10 and was pacing above 87 trillion the following week.
Token volume is not the same as useful work, as the speaker acknowledged.
The practical piece is routing. A service can select models based on complexity, price, speed, and reliability. Cheap work can go to a fast model, while difficult work can go to a frontier model.
Stripe demonstrated an agent deploying a service with hosting, a database, a sandbox, credentials, usage tracking, and billing. The service then charged other agents $2 for a review.
The proposed stack connects model costs with usage measurement, payments, fraud protection, and revenue. It also includes stablecoin tools and the ability to store funds.
The source argues that small teams can rent capabilities that once required departments, integrations, vendor negotiations, capital, and months of work. The stack does not provide a worthwhile problem, judgment, taste, trust, or customers.
Stripe’s customer data is only an early signal. Its 5 million customers are online, likely to grow quickly, and disproportionately connected to San Francisco. The speaker did not treat them as the whole economy.
Apple’s Cuts Are Clear. The Explanation Is Not.
Apple cut 200 jobs: 100 from the Vision Pro group and 100 from Siri and its AI software teams.
The speaker said the cuts happened because:
“the work stopped existing.”
His argument was that older expertise did not fit the new architecture. That explanation was not independently established.
He also said Apple reportedly pays $1 billion a year for a custom model, although the surrounding architecture reference was unclear.
His larger argument was that models are becoming commodities, while Apple’s stronger assets are its two billion devices and access to users’ location, health, and messages. He also cited a 67% price cut without explaining it further.
The cuts are established. The architectural explanation remains the speaker’s claim.
Faster Tools Still Produce New Failure Modes
Codex CLI 0.149.1 added thread-source classification, image budget enforcement, and memory consolidation tagging. Its startup was also reported as roughly 25 times faster after Charlie Marsh reworked the lifecycle.
At the same time, the unsupported prompt_cache_retention parameter reportedly caused 400 errors in Codex Desktop with GPT-5.6 Sol after tool calls. The stated workaround was to use the CLI directly.
Other reports covered a Claude Code launch segfault and Cursor Cloud Agents friction involving fragmented environments, web auto-mode, and media handling.
The same daily note collected prompting guides, scoring workflows, agent toolkits, skills, job-search automation, an open-source program, and community claims. Representative examples included:
- A toolkit listing 30 agents, 64 skills, 33 commands, and security tests
- A collection of 132 hand-curated
SKILL.mdplaybooks - A local proxy with approximately 46k stars, claiming roughly 1.3B free tokens per month across 48 providers
The first Cursor Hackathon in Mar del Plata featured WhatsApp bots, farm-management applications, and custom Cursor plugins.
The tools are becoming faster and richer. The reported failures show that the systems around the model can still be fragile.
Start with your best work and the thinking process behind it. Define success, then ask which model accelerates that loop.
The model is only one part of the result.