GLM-5.3 in Claude Code 2.1.280
skills off
This is the unfinished output at the moment the run ended (Harness + model mismatch). See the review below.
Wall time 8m 13sCost $0.23 estimated
Review
Calm-but-moody night scene: a moon column on a banded sea, a strong beam through haze, a keeper's hut and surf at the cliff foot.
The full review
Calm-but-moody night scene: a moon column on a banded sea, a strong beam through haze, a keeper's hut and surf at the cliff foot. The tower is drawn slightly twisted and uneven, the rain is very faint, and a white rope-like line runs along the cliff edge with glitchy surf blobs.
Verdict by the AI judge (provisional).
Harness + model mismatch. Reclassified 2026-09-24 (published as ok): the run ended on 'API Error: 400 This model does not support image inputs' when GLM-5.3 read its own render (…/render1.png) with Claude Code's Read tool; the artifact was already written but the self-review was cut short (Claude Code x text-only model). Re-run on the text-only relay: lighthouse_night__claude-code__glm-5.3__attempt2.
Attempt 1 of 2
Every try of this model in this harness on this prompt, oldest first. A later attempt is the cell's result: attempt 2.
- Attempt 1 Harness + model mismatch this run8m 13s$0.232026-09-24
- Attempt 2 OK the cell's result16m 19s$1.492026-09-25
Bugs
- twisted uneven tower
- very faint rain
- stray white line along the cliff edge
Scores
Not scored: only runs that finished on their own are scored.
Downloads
Files actually published for this run. Original output, site-made previews and post-run conversions are labelled separately.
The exact prompt
agentic harness (save the file; self-render allowed), 378 characters. Every result for this prompt
Create an SVG of a lighthouse on a rocky headland at night: stormy sea, the lighthouse beam cutting through rain and mist, moonlight on the waves, as atmospheric and painterly as you can. Save it as lighthouse.svg in the current directory. You may render and inspect your own output (headless chromium is available at /usr/bin/chromium) and iterate on your own before finishing.
How it was made
Step by step
The prompt, every agent turn and tool call (tool name and a one-line summary; inputs and outputs are not shown), then the final message. The full machine-format transcript is published as a scrubbed download.
Show all 9 steps
Loading the timeline…
Download the full transcript Claude Code session JSONL, 274 KB. Scrubbed: anything that looks like a key, token, e-mail address or phone number is replaced, and account details and the test machine's file paths are removed.
Cost breakdown
Headline cost $0.23, estimated from the token counts at list price.
| Tokens by type | Tokens | $ per 1M | Cost |
|---|---|---|---|
| Input (uncached) | 48,253 | $1.40 | $0.0676 |
| Cache write | 0 | $0.00 | $0 |
| Cache read | 112,895 | $0.26 | $0.0294 |
| Output | 30,254 | $4.40 | $0.1331 |
| of which reasoning | not recorded | billed as output |
- Total computed from tokens
- $0.2300
- Total reported by the harness
- none
Tokens summed from the Claude Code session JSONL (per message id, incl. subagent files); if Claude Code logged zero usage for this unrecognised model, the /cost screen's per-model usage is used instead. Claude Code tags glm-5p3 as unrecognized_model and its own $ figure has costBasis unknown, so it is NOT used: cost = tokens x Fireworks GLM-5.3 price.
Prices: Fireworks GLM-5.3 serverless price (STATE.md 22:42Z; same as the OpenCode/models.dev catalog entry fireworks-ai/glm-5p3), 2026-09-23.
Run notes
- Mode: interactive TUI in tmux (160x48), inside bubblewrap sandbox
- Isolation: bwrap: private /tmp, HOME and TMPDIR inside run dir, own CLAUDE_CONFIG_DIR copy; the project files, other runs, /tmp/hx_* not mounted

