Today we’re announcing Hoshi 1.0, Shogo’s own agent model — and the results that come with it. Across three independent, harness-neutral evaluations, Shogo powered by Hoshi 1.0 beats OpenClaw running Claude Haiku 4.5 — and OpenClaw running OpenAI’s gpt-5.4-nano — on every single one — agentic tool use, deterministic structured output, and multi-language coding. And on the coding benchmark, Hoshi 1.0 reaches a perfect score while running an order of magnitude cheaper than Haiku.
Hoshi 1.0 is the model that runs inside Shogo’s agents. It was built and tuned for exactly the work our users do every day: driving real tools, editing real files, and producing clean, parseable output the first time.
The headline: a clean sweep
| Eval | Shogo + Hoshi 1.0 | OpenClaw + Haiku 4.5 | OpenClaw + gpt-5.4-nano |
|---|---|---|---|
| ClawBench Core v1 (agentic, overall 0–1) | 0.844 | 0.797 | 0.782 |
| ClawEval (1,220 checkpoints)* | 88.4% | 84.7% | 76.1% |
| Aider Polyglot (83 tasks, pass@2) | 83 / 83 (100%) | 81 / 83 (97.6%) | 76 / 83 (91.6%) |
| Aider — model cost | $0.41 | ~$5.2 | ~$1.36 |
Three evals, three wins — each one measuring a different muscle: tool-use and trajectory (ClawBench), deterministic structured output (ClawEval), and multi-language coding (Aider Polyglot). We’ve also added a third stack — OpenClaw running OpenAI’s gpt-5.4-nano — and Hoshi 1.0 stays ahead of it across the board, too.
Hoshi 1.0 vs Claude Haiku 4.5, benchmark by benchmark
| Benchmark | What it measures | Hoshi 1.0 | Claude Haiku 4.5 | gpt-5.4-nano |
|---|---|---|---|---|
| ClawEval* | Deterministic structured output | 88.4% | 84.7% | 76.1% |
| ClawBench Core v1 | Agentic tool use | 0.844 | 0.797 | 0.782 |
| Aider Polyglot | Multi-language coding, 83 tasks | 83 / 83 | 81 / 83 | 76 / 83 |
Head-to-head on each benchmark, Hoshi 1.0 leads — and on the deterministic ClawEval suite, Shogo’s full agent tops the v2 leaderboard at 88.4%.
Hoshi 1.0 on the major public benchmarks
Our internal evals aren’t the only place Hoshi 1.0 separates from both Claude Haiku 4.5 and OpenAI’s gpt-5.4-nano. On the industry-standard benchmarks the rest of the field reports, Hoshi 1.0 posts frontier-tier agentic numbers — and on Terminal-Bench Hard, the toughest standardized test of autonomous terminal work, it leads both alternatives by double digits.
| Benchmark | What it measures | Hoshi 1.0 | Claude Haiku 4.5 | gpt-5.4-nano |
|---|---|---|---|---|
| Terminal-Bench Hard | Autonomous terminal / agentic tasks | 43.2% | 27.3% | 33.3% |
| SWE-bench Pro | Hard, multi-file real GitHub issues | 57.2% | n/r | 52.4% |
| ClawEval (Pass³) | Reliability across 3 agentic runs | 64% | n/r | n/r |
Terminal-Bench Hard is the head-to-head highlight: 43.2% for Hoshi 1.0 vs 33.3% for gpt-5.4-nano and 27.3% for Haiku 4.5 — a double-digit lead over both on exactly the kind of long-horizon, tool-driven work Shogo agents do all day. On SWE-bench Pro — the harder, multi-file cut of real GitHub issues — Hoshi 1.0 resolves 57.2%, ahead of gpt-5.4-nano’s 52.4% and squarely in frontier territory for agentic software engineering, while sustaining 64% Pass³ reliability on ClawEval. (“n/r” = not publicly reported on that specific cut.)
Coding: a perfect score
On the full 83-task Aider Polyglot set, Hoshi 1.0 solves every exercise across 49 JavaScript and 34 Python tasks on the first attempt, while Haiku misses two. Hoshi is also faster end-to-end — roughly 37s per task versus Haiku’s ~45s.
| Shogo + Hoshi 1.0 | OpenClaw + Haiku 4.5 | OpenClaw + gpt-5.4-nano | |
|---|---|---|---|
| JavaScript (49) | 49 / 49 | 48 / 49 | 45 / 49 |
| Python (34) | 34 / 34 | 33 / 34 | 31 / 34 |
| Total | 83 / 83 (100%) | 81 / 83 (97.6%) | 76 / 83 (91.6%) |
Frontier results at a fraction of the cost
The most striking part isn’t just that Hoshi 1.0 wins — it’s how cheaply it does so. Running the identical 83-task coding set, Shogo + Hoshi 1.0 cost $0.41 total versus ~$5.2 for OpenClaw + Haiku 4.5 (and ~$1.36 for OpenClaw + gpt-5.4-nano) — an order of magnitude cheaper than Haiku for an equal-or-better score.
| Shogo + Hoshi 1.0 | OpenClaw + Haiku 4.5 | OpenClaw + gpt-5.4-nano | |
|---|---|---|---|
| Total cost | $0.41 | ~$5.2 | ~$1.36 |
| Per task | $0.005 | ~$0.06 | ~$0.016 |
That gap comes straight from the token economics. Per million tokens:
| Hoshi 1.0 | Claude Haiku 4.5 | gpt-5.4-nano | |
|---|---|---|---|
| Input | $0.15 | $1.00 | $0.20 |
| Cached input | $0.01 | $0.10 | $0.02 |
| Output | $0.30 | $5.00 | $1.25 |
Hoshi 1.0 is about 7× cheaper on input and ~17× cheaper on output than Haiku 4.5 — and output is the dominant cost axis for code generation.
Agentic tool use
On ClawBench Core v1 — 19 agentic tasks run three times each with native tools — Hoshi 1.0 leads overall at 0.844, ahead of Haiku 4.5 at 0.797 and gpt-5.4-nano at 0.782, with the biggest margins on the dimensions that matter most for getting work done:
| Axis | Shogo + Hoshi 1.0 | OpenClaw + Haiku 4.5 | OpenClaw + gpt-5.4-nano |
|---|---|---|---|
| Overall | 0.844 | 0.797 | 0.782 |
| Completion | 0.898 | 0.816 | n/r |
| Trajectory | 0.757 | 0.721 | n/r |
Hoshi 1.0’s standout wins came on test-writing and hallucination resistance — tasks where it stayed grounded and finished the job while the competing stack stalled.
Structured output
On ClawEval — 59 roles and 1,220 deterministic checkpoints — Shogo’s full agent on Hoshi 1.0 scored 88.4%*, #1 on the ClawEval v2 leaderboard and ahead of both Haiku 4.5 (84.7%) and every prior baseline. It’s a clean confirmation that Hoshi 1.0 produces correct, parseable output under strict, automated grading.
From the team
“From day one, we built Shogo on a different premise than the rest of the industry: the future of software should be self-evolving, model-agnostic, and truly open — not locked into a single lab’s roadmap or ecosystem,” said Guru Angisetty, CEO and Co-Founder of Shogo. “Hoshi 1.0 is a major step toward that future. It gives Shogo its own agent-native intelligence layer while preserving the freedom to use the best models across the market. The fact that Hoshi 1.0 is outperforming frontier alternatives like Claude Haiku 4.5 and GPT Nano 5.4 across our evals, while running at roughly a tenth of the cost, proves that open AI platforms do not have to compromise on performance, economics, or control. That shift is what democratizes capable agents — putting them in front of every team, not just the ones with frontier-lab budgets.”
“Shogo is built to be the most powerful place to create AI agents that evolve — agents that turn your workflows into living apps and systems, live inside them, and keep improving them as your business changes,” added Russell LaCour, CTO and Co-Founder. “Hoshi 1.0 is the intelligence layer that makes that real. It drives the tools, reshapes the workflows, and returns clean, reliable results — at frontier-tier performance and a fraction of the cost. That’s what lets every team build software that learns, adapts, and improves itself, instead of waiting on someone else’s roadmap.”
About Shogo
Shogo is the platform where AI agents build the software you need instead of replying with text. Describe what you want, and a Shogo agent — now powered by Hoshi 1.0 — builds it: a live dashboard, an automated workflow, a visual interface wired to your real tools. Deploy your first agent in minutes on the free tier, or browse the templates to see what teams are building today.
* ClawEval reflects Shogo’s full agent on the ClawEval v2 leaderboard (Phase H, 1,220 checkpoints), scored with the lenient (uniform) scorer: 1,078 / 1,220 = 88.4%, #1 on the leaderboard. On the stricter published scorer, Shogo ties for #1 at 86.9%. Figures shown for other models are their own published results.