Together Compute’s DeepSWE routing results indicate GLM-5.3 solves a higher share of problems per dollar than Fable 5: 87.6% solved for roughly $16 versus 69.7% for $21.63. The GLM figure is aggregated from four attempts rather than a single run, so it requires cautious interpretation. The same day the Financial Times reported that Anthropic’s most powerful models are losing ground to cheaper alternatives.
What happened and why it matters
Together Compute’s routing math highlights that when systems can retry, cost-efficiency—how many passing runs you get per $100—can matter more than which model wins a single run. Their numbers place GLM-5.3 at 17 solves per $100 on DeepSWE, compared with 3 solves per $100 for Fable 5. For agent architectures that can retry, cheaper models with higher aggregate pass rates can be more economical than a pricier single-shot winner.
A key caveat: the 87.6% figure for GLM-5.3 comes from four attempts aggregated, not a one-shot success rate. That means results depend heavily on retry strategy and routing logic and shouldn’t be treated as a direct one-to-one comparison without replicating the exact setup.
Other industry updates (brief items)
- Empero now offers a free OpenAI-compatible community endpoint for Qwen3.8-27B-FP8; it’s zero-setup to point a client at that 27B model. No uptime, rate-limit, or data-handling guarantees were specified, so it’s suited for test harnesses for now.
- Archal launched API sandboxes for AI agents with Slack, Linear, Datadog and 20+ stateful environments for CI and evaluation — addressing a part of agent testing teams often fake with mocks.
- terminal-code runs VS Code inside the terminal by combining code-server with a terminal-browser, keeping editor, diffs, SSH sessions and review flow in one pane.
- Bezalel bundles memory, email, texting, payments, a computer, sandboxes and connectors behind a single MCP URL to let you test capable agents without wiring multiple integrations; it remains alpha and connector claims vary.
- SenseNova released U1.5-8B-MoT on Hugging Face, an 8B any-to-any multimodal model suitable for local routing tests, though the model card lacks benchmarks.
- Audio8 published TTS-Preview-0.1B, billed as a very small zero-shot TTS model (0.1B parameters), enabling local per-request voice synthesis tests.
Security, tools and research notes
- Anthropic placed Mythos 5, its non-public model tier, behind Claude Security for scanning GitHub repositories and proposing patches — indicating the strongest code/security model is being used in an automated scanning context rather than general chat.
- Fabien Sanglard published his agent.md file used to hold LLM-assisted code to his quality standard; it’s a copyable instruction file worth diffing against your own.
- The Model Context Protocol maintainers added progressive tool discovery and a standard tool-result contract to the roadmap to reduce the problem of loading hundreds of tools into context and to give clients a single result shape to build against.
- Google Cloud published five patterns for long-horizon agents: stable prefixes, background learning, persistent workspaces, explicit failures and guard chains; these target quietly accumulating bugs such as cache burn, memory loss, or mis-marked sub-agents.
Fine-tuning, rate-limits and performance numbers
- A Gemma 4 12B model was fine-tuned for tool-calling and reported a 2.7× improvement; the model was chosen in part because it fits comfortably in 16 GB of VRAM, suggesting tool reliability can be improved on commonly available hardware.
- OpenAI Codex lead Tibo highlighted two drivers of rate-limit pain: image-heavy long sessions with repeated compactions, and Computer Use getting expensive at the p95+ tail. The practical advice is to avoid embedding images in runs you expect to compact.
- Reported throughput numbers: GLM-5.2 753B ran at 14.9 tokens/s on a 96 GB GPU; DeepSeek-V4-Flash 284B at 22 tok/s on 32 GB; Qwen3.6-35B at 39.3 tok/s on an 8 GB card. If reproducible, the "too big to serve locally" threshold has moved down, and the 8 GB Qwen result is broadly accessible to many developers.
Clone & Run and open-source tooling highlights
- Apache Maka: a local-first agent workspace that logs model messages, tool calls, permissions, and termination events as first-class records; still in Apache incubation.
- Paseo: an open-source control surface for Claude Code, Codex, Copilot and OpenCode across local machines and VPS.
- Anthropic’s claude-plugins-community: a read-only mirror of the Claude Code and Claude Cowork plugin directory, showing what’s actually in the marketplace.
- OpenHuman: an open-source personal AI with local-first memory and agent-fleet orchestration combined.
- agent-safe-pipeline: a reference architecture where agents can propose actions but cannot approve them, sending final decisions to an independent policy layer.
- open-slide: turns a prompt into an interactive deck by having coding agents write React components so the generated artifact remains editable by the agent.
- Ruflo, oh-my-subagents and others provide larger implementations for multi-agent coordination, subagent persistence and auditing.
Usage statistics
- a16z reports that since February the fastest-growing Codex adopters are outside tech: legal up 108×, sales and recruiting 41×, marketing 26×, healthcare 24× — suggesting non-engineering users are converting repeatable knowledge work into Codex-shaped tasks.
Conclusion
Together Compute’s routing results underline that routing strategies and ability to retry can change which models are most cost-effective for coding workloads. Cheaper or locally runnable models may outperform pricier single-shot models in aggregate cost-per-pass metrics. Because the GLM-5.3 figure aggregates multiple attempts, teams should run their own routing experiments with real prompts and infrastructure before switching defaults.
