Four days ago, Anthropic dropped Claude Opus 5.5 into Claude Code and cut the price of flagship-tier agentic coding by 40% in one release. The same week, OpenCode crossed 208,000 GitHub stars, 950 contributors, and 16 million monthly developers
- making it the most-starred open-source coding agent by a margin that's stopped being close. Two things shifted simultaneously, and together they force a decision most engineering teams have been deferring: which agent stack do you actually standardize on?
What Opus 5.5 actually changes for coding agents
Anthropic released Claude Opus 5.5 on September 22, 2026. It performs at the level of Fable 5.1 on most work and, at default settings, costs 40% less to run than Opus 5. Input and output tokens are priced at $4 and $20 per million, a 20% cut from Opus 5. The cache-read cut is steeper: cache reads fell from $0.50 to $0.20 per million tokens - a 60% cut - because Opus 5.5 prices cache hits at 5% of the input rate, where other Claude models use 10%.
The 40% headline is worth reading carefully. Anthropic says it costs 40% less on typical workloads at default effort because it also uses fewer tokens per task; that second part depends on your workload and is not yet independently confirmed. A fixed-token back-of-napkin test gives a different number: hold usage fixed and the arithmetic is about 31% less, before any change in task length or success rate. The honest answer is that your actual saving depends heavily on cache hit rate and output length. Run your own workload through both models for a week before trusting the rate card.
What's cleaner is the benchmark picture. At medium effort - the new default - Opus 5.5 matches or beats GPT-6 Astra at max effort on FrontierCode and GDPval-AA for roughly a fifth of the cost per task. Fable 5.1 no longer has a performance case for everyday coding at 2.5 times the price.
One operational thing teams are missing: as of Claude Code version 2.1.280, released September 22, Pro plans got moved from Sonnet to Opus at the same time. If your team is on Team Standard and you have token budgets set for Sonnet-grade spend, those budgets are now wrong. Check your billing settings before the next sprint.
Where open-source coding agents now stand
By GitHub stars as of September 22, 2026: OpenCode (209,405, MIT), OpenAI Codex CLI (125,962, Apache-2.0), Gemini CLI (107,130, Apache-2.0), Cline (69,074, Apache-2.0), Goose (54,568, Apache-2.0), Aider (49,118, Apache-2.0), and Kilo Code (27,390, MIT).
OpenCode's lead is not just cosmetic. It is model-agnostic, connecting to 75+ providers; editor-agnostic across terminal, desktop app, VS Code extension, or web UI; privacy-first with zero data retention; and no training on your code, with support for air-gapped environments.
It runs locally and stores conversations in SQLite, giving you full control over which models process your code.
The thing most comparisons miss: OpenCode isn't really competing with Claude Code on benchmark scores. It's competing on control surface. Claude Code provides deep semantic understanding of code; OpenCode treats code files the same way it treats any other file - as text to be read and modified, without inherent awareness of language semantics or test coverage. For a team that wants to write and refactor code autonomously, Claude Code's harness optimization matters. For a team that runs a range of model-dependent tasks (some cheap, some heavy) or operates under data-residency rules, OpenCode's model-switching is the feature.
One genuinely non-obvious consequence of OpenCode's position: when Anthropic ships Opus 5.5, OpenCode users can call it immediately through their existing API key.
The install command is a single curl, and you run a specific model variant with opencode run --model opencode/claude-opus-5-5.
Closed-agent users get the model through the vendor's harness on the vendor's schedule. The distinction matters less when models are stable, and a lot when the frontier moves every two months - which is the pace right now.
Opus 5 held the flagship slot for just two months after its July 24, 2026 release. The pace matters for anyone pinning model versions in production: the model you standardize on this quarter may have a cheaper sibling before the rollout finishes.
How the two stacks actually compare
The honest comparison is not Claude Code vs OpenCode - it's "closed harness, best-in-class model optimization" vs "open harness, model portability." Both are valid; the right answer depends on what your team values.
| Dimension | Claude Code | OpenCode |
|---|---|---|
| Model access | Anthropic models only | 75+ providers + local via Ollama |
| Benchmark score (Terminal-Bench 4.0) | 57.9% (Fable 5.1) | Depends on model you route |
| Cost per task | Lower after Opus 5.5 price cut | Pay only for tokens; no tool fee |
| Data residency | Anthropic-controlled | Self-hosted, air-gap capable |
| Code semantics | Deep (LSP + harness optimization) | Text-level; no native LSP |
| License | Proprietary | MIT |
| Breaking changes on model upgrade | Managed by Anthropic | You own the upgrade path |
A common production stack: Claude Code or Codex for heavy agent work, Copilot or Cursor for inline completions, and one open-source agent (OpenCode, Cline, or Kilo Code) for model flexibility. Most teams end up running two tools, not picking one winner.
Beagle in action#engineering, 10:02amOne thing the leaderboard doesn't tell you
On Terminal-Bench 4.0 the jump from xhigh to max effort added 0.3 points for Codex + GPT-6 Astra while raising cost per task 28-39%. Every agent above is a harness around a model. The same 57.9% score cost $7.12 per task on one pairing and $14.76 on another.
That's the thing worth sitting with. The benchmark score you read on a leaderboard is the headline. The per-task cost at the effort level your team actually runs is the number that shows up in your AWS bill. Neither Claude Code nor OpenCode will tell you that number upfront; you have to instrument it yourself.
A teammate like Beagle can pull this directly into Slack - log the model, the effort level, and the token count per merged PR for a sprint, and the cost-per-task curve becomes something your team can actually act on rather than estimate.
The open-source leaderboard is interesting context. The model release is the thing that changes your budget this week.
Open source coding agents: common questions
What is the most popular open-source coding agent right now?
OpenCode leads by GitHub stars with over 208,000 stars, 950 contributors, and 16 million monthly active developers. It is MIT-licensed, model-agnostic across 75+ providers, and runs in terminal, desktop, or IDE. For IDE-first teams, Cline and Kilo Code are the strongest alternatives.
Does Claude Code work with models other than Claude?
No. OpenCode supports 75+ model providers versus Claude Code's Anthropic-only approach. If you want to run DeepSeek, GPT-6 Sol, or a local model through Ollama in the same agent harness, you need an open-source alternative like OpenCode or Cline.
How much cheaper is Claude Opus 5.5 than Opus 5?
Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5 and $25, with cache reads cut from $0.50 to $0.20. Anthropic claims 40% lower running costs on typical workloads, which combines the token price cut with fewer tokens consumed per task. The fixed-token saving is closer to 31%.
What broke when Opus 5.5 shipped in Claude Code?
The model ID is claude-opus-5-5, the 1M-token context window and 128K max output carry over from Opus 5, and thinking is now always on. Four documented breaking changes affect code already running on Opus 5. Anthropic's release notes list them; review before upgrading any production integration.
Should a small engineering team use an open-source or closed coding agent?
The practical split: closed agents (Claude Code, Codex) give you better benchmark performance and a managed model upgrade path. Teams that can't route sensitive data through external APIs - healthcare, finance, legal, government - find that self-hosting removes the data-residency problem at the architecture level rather than the contract level. Start with whichever your team already has API access to, instrument cost-per-task after one sprint, then decide.