What happened
Startup Fortune reports that ForgeCode, an open-source coding harness from Tailcall, scored 81.8% on the self-run TermBench 2.0 leaderboard. The article says ForgeCode submitted its own result, making the leaderboard unsuitable as an independent comparison with Claude Code or Codex. On the Princeton- and University of Chicago-run SWE-bench Verified , Startup Fortune reports that ForgeCode paired with Claude 4 scored 72.7%, close to leading Claude Code and Codex configurations. The source also reports that ForgeCode’s orchestration layer raised Gemini 3.1 Pro’s TermBench score from 55% to 80.2%.
Startup Fortune reports that ForgeCode leads TermBench 2.0 with an 81.8% score and that the company submitted the result to the leaderboard. Because the is described as self-run and shaped by ForgeCode, the article treats the result as a company-controlled comparison rather than independent evidence that ForgeCode beats Claude Code or Codex.
The article reports a narrower result on SWE-bench Verified: ForgeCode paired with Claude 4 scored 72.7%, within a few percentage points of top Claude Code and Codex configurations. The source does not provide the exact comparison scores in the supplied text, and these claims have not been independently confirmed here.
Startup Fortune also reports that Tailcall increased Gemini 3.1 Pro’s TermBench 2.0 score from 55% to 80.2% by using ForgeCode’s orchestration layer. It cites a separate Scale AI comparison in which three systems using identical Claude Opus 4.5 weights scored between 50.2% and 55.4% on SWE-bench Pro. These figures are reported by Startup Fortune and are not independently validated in this assessment.
Source details: startupfortune.com ↗
Why it matters
The report highlights how much agent scaffolding, design, retry behavior, and context management can influence coding- results. A higher score may reflect optimization around a particular test rather than a more capable underlying model. For developers choosing coding tools, benchmark governance and reproducibility may therefore matter as much as the headline result. Startup Fortune also reports no evidence that production teams are replacing Claude Code or Codex with ForgeCode, leaving real-world adoption unconfirmed.
The reported results suggest that the software surrounding a model can materially affect coding-agent performance. Planning, context handling, retries, and test-specific tuning can change scores without changing the model’s weights.
That makes self-hosted or narrowly defined leaderboards weak evidence for broad claims about model or product superiority. The source’s independently validated comparison is more relevant, but it still does not establish which tool is best for a particular developer or team.
The practical implication is to evaluate coding assistants with multiple benchmarks and representative repository tasks, while checking security and reliability. Startup Fortune reports no evidence of meaningful production migration from Claude Code or Codex to ForgeCode, so adoption remains unknown.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
Independent reruns using identical models, prompts, budgets, and retry limits would show whether ForgeCode’s advantage generalizes beyond its own leaderboard. Developers should also watch for evidence from sustained production use, including reliability, security, workflow retention, and maintenance costs. The source describes ForgeCode as model-agnostic and compatible with Bash, Zsh, and Fish, but it does not establish pricing, support terms, general availability, or whether ordinary users can reproduce the reported scores.
Look for independent reruns that disclose prompts, model versions, context limits, retry budgets, infrastructure, and evaluation procedures.
Watch for evidence from real development teams rather than repository stars and forks alone. Startup Fortune notes that ForgeCode’s reported GitHub activity indicates engagement, but not displaced workflows.
Pricing, access, support, and reproducibility remain undocumented in the supplied source. It is also unknown whether the reported performance persists across repositories, languages, and security-sensitive production work.