EvalSignal 010 / Independent model evaluation
Why OpenCode may be the next AI acquisition
This one isn't a benchmark. It's a bet.
We spend most of these issues comparing models. This issue is about the thing developers put in front of the models — and why we think it is more acquirable than most people realize.
OpenCode is already the de facto terminal agent
OpenCode is not a scrappy alternative anymore. By its own numbers it has crossed 195,000 GitHub stars, 950 contributors, and more than 16 million monthly developers. Its GitHub repository has since passed 200,000 stars. Multiple 2026 comparisons of terminal coding agents now call it the dominant open-source option in the category, ahead of Gemini CLI, OpenHands, Aider, and Goose.
It talks to more than 75 model providers out of the box: Anthropic, OpenAI, Gemini, Bedrock, local models through Ollama, and anything OpenRouter can reach.
That breadth is the whole thesis. OpenCode does not bet on one lab's model being best forever. It bets that someone's model will be best at any given moment, and it wants to be the terminal you are already sitting in when that happens.
It is everywhere a developer already is
You do not create an account or take a subscription just to install it. You run one install line, or reach for whatever is already on your machine: Homebrew on macOS and Linux, pacman or AUR on Arch, Nix, Scoop and Chocolatey on Windows, or plain npm i -g.
On a locked-down box where you cannot install much of anything else, there is usually a path to OpenCode. That ubiquity is a distribution moat competitors cannot easily buy their way around. It took two years of default-open, no-lock-in positioning to earn.
The growth curve backs it up: OpenCode gained roughly 18,000 stars in two weeks during a single stretch of early 2026, briefly outpacing Claude Code's total star count in a single day. That kind of alpha-stage spike, sustained rather than one-off, is the sort of organic growth signal acquirers pay premiums for.
People run it alongside Claude Code and Codex
This is the part that should worry the labs more than it comforts them. OpenCode is not necessarily stealing developers away from Claude Code or Codex. It sits next to them as the neutral layer developers reach for when they want to switch models mid-task, compare outputs, or avoid being locked into one vendor's usage limits.
Community rankings of CLI agents frequently place OpenCode as the orchestration layer other tools plug into, precisely because it does not insist on a single backend.
Anthropic already tried to make it go away
In January 2026, Anthropic moved to block OpenCode and several other third-party harnesses from using Claude Pro and Max OAuth tokens instead of metered API keys. It escalated through the spring: a February terms rewrite, a formal legal request that forced OpenCode's maintainers to strip Claude-specific code in March, and an April cutoff that briefly took OpenCode off Claude entirely.
OpenAI stepped into the gap, bringing ChatGPT Plus and Codex access into the tool. By June, Anthropic had reversed course into a capped, API-billed credit pool rather than a full ban.
Read that sequence again: a frontier lab spent five months trying to contain a single open-source terminal tool, failed to kill it, and watched a rival use the standoff to get closer to it instead.
That is not the profile of a nuisance project. It is the profile of critical infrastructure — the kind labs eventually decide is safer to own than to fight.
Everyone in this market is short on distribution
Look at what large AI players have actually been buying:
- Nvidia agreed to buy Hugging Face for roughly $12.9 billion, on top of a reported $6 billion licensing deal with Poolside for open-model development.
- Stripe acquired OpenRouter, the routing layer between developers and dozens of model providers. The price was not disclosed publicly; reports placed it around $7.5–8 billion.
- Google and Meta both ship or develop coding agents, but neither has OpenCode's contributor base or model-neutral positioning.
- Chinese labs such as DeepSeek, Z.ai, and Moonshot are racing to get their models into provider-agnostic tooling that can win developer mindshare outside their home market.
The company behind OpenCode is Anomaly, the team also known for SST, OpenNext, and OpenAuth. Y Combinator lists the company at only 24 people. That is a small organization sitting underneath one of the most-used pieces of AI developer tooling in existence.
Small team, massive distribution: that combination does not usually last long once the rest of the market notices it.
Who would actually want it, and why?
- OpenAI already has the working relationship and would be buying goodwill along with distribution.
- Anthropic has the most complicated history but also the most to gain from controlling, rather than fighting, the layer where developers decide which model to call next.
- Meta and Google have the least credible open-agent story of the major labs and the deepest pockets to fix that with one purchase instead of years of catch-up.
- A Chinese lab or Nvidia-style infrastructure buyer gets an instant, trusted distribution channel into millions of Western developer terminals.
- SpaceXAI, the combined Musk entity, has shown it is willing to write very large, fast checks to fill gaps in its AI stack. A credible coding agent is one of the more obvious gaps left.
Bottom line
The bigger picture is that a war is coming for every layer of the agent stack — planners, routers, browser agents, terminal agents, all of it — as labs and infrastructure players realize distribution matters as much as model quality.
OpenCode did not start that war, but with more than 200,000 stars, 16 million monthly developers, and a position no single lab controls, it is currently leading it. It has already survived one major lab trying to shut it out and emerged with a broader provider story.
In a season where a chipmaker agreed to pay $12.9 billion for a model hub and a payments company paid billions for a routing layer, a terminal agent used by 16 million developers a month is not an outlier to watch. It may be next in line.
Do not be surprised if you see it.
This is opinion and speculation based on public reporting, not investment advice. We hold no position in and have no inside knowledge of any company named above. Do your own research before acting on any of it.
Bonus benchmark
Does DeepSeek V4 Flash Vision Exp actually see better than Qwen and GLM?
Short answer: no.
We put DeepSeek's new experimental multimodal route, deepseek/deepseek-v4-flash-vision-exp, through the same 200-case harness we used on Qwen3.8 Flash and GLM-5.3 Flash: 100 text-planning cases and 100 production vision cases.
| Model | Pass | Rubric | Median |
|---|---|---|---|
| Qwen3.8 Flash | 79/100 | 94.8% | 8.42s |
| GLM-5.3 Flash | 76/100 | 93.8% | 5.49s |
| DeepSeek V4 Flash Vision Exp | 68/100 | 89.5% | 5.13s |
DeepSeek's eyes are the fastest of the three and the weakest: eleven strict passes behind Qwen and eight behind GLM. The gap widens most on hard screenshots. It holds up on toasts, tables, and kanban boards, but drops to 2/5 on authentication and calendar screens and 0/5 on modal overlays, a category all three models failed.
Where DeepSeek wins outright is text planning: 45 ideal-name-or-terminal outcomes versus 41 for GLM and 36 for Qwen, with the tightest p95 latency of the three. That tracks with its 0731 checkpoint's re-post-training for agentic tool use and a newly attached speculative-decoding module.
Verdict: DeepSeek V4 Flash Vision Exp is the best planner of the three Flash-tier models and a serviceable, not superior, vision model. If your workload is mostly text and tool calls with occasional screenshots, its eyes are good enough. If browser vision is a first-class workload, Qwen3.8 Flash still leads.
