Video Summary - GLM 5.3 Flash & DeepSeek Harness Slash AI Costs

Key technological/product concepts

Cost differences for “same model, different size/variant” (GLM 5.3)
- GLM 5.3 Flash: ~$0.07–0.075 per 1M input words (discounted pricing mentioned with the note that exact pricing was discussed with dates).
- Larger GLM 5.3: about $1.40 per 1M input words.
- Claim: Flash can be ~20x cheaper for input, while still delivering “shockingly good” performance for its parameter scale.
DeepSeek Harness as an open-source “agent harness”
- MIT licensed → the harness itself costs nothing.
- The expensive part of an agent is model inference (“the model thinking behind it”).
- Switching models/providers is described as configurable via four fields:
  - provider name
  - web address / endpoint URL
  - protocol
  - API key
- The video emphasizes the harness can run coding agents, using different models for easy vs. hard tasks.

DeepSeek Harness model configuration approach
- Swap models by filling required connection fields:
  - provider
  - endpoint URL
  - protocol
  - key
Host compatibility “toggling” (gateway quirks)
- The harness includes switches for:
  - whether system instructions can use an unknown role
  - how to set field caps for reply length
- Warning: these toggles don’t verify the endpoint; they only describe expected behavior. Wrong settings can cause silent failures.
Plugin ecosystem
- The harness supports plugins (no “privileged core” stated), enabling builders to replace:
  - model adapters
  - tool lists
  - even agent loop logic
- Example plugin: FlyKit
  - fetches the live model list on startup instead of using a frozen built-in list

Model capability comparisons (Flash vs larger GLM 5.3)
- “General intelligence” style score: Flash ~42, ranked 4th out of 112.
- Coding performance score: ~71.5, described as beating a large portion of models.
- “Agentic index” (planning/following a plan): Flash ~51.5, beating ~90% of models.
- Main takeaway: despite being much cheaper, Flash is near the top on coding/agent-relevant tests.
Agent output quality isn’t the main cost driver
- For agent coding loops, the video stresses cost is dominated by how much input is resent, not by how much the model outputs.

Agent “input vs output” ratio (OpenRouter measurements)
- Over ~10 days, users reportedly sent:
  - ~1.25 trillion words in to GLM 5.3 Flash
  - ~32 billion words out
- Implies roughly 38 input words per 1 output word.
- Reason given: the harness re-reads/scans files repeatedly each step, so the model receives large contexts repeatedly.
Caching changes the effective input price
- Effective cost depends on cache hit rate by host.
- Cache hit rates range ~27% to 90%, depending on host.
- Weighted averages from OpenRouter are used to estimate real buyer cost.
Sticker prices can be misleading
- Promotional input/output pricing differs from real averages:
  - Input cheaper than advertised in practice
  - Output more expensive than advertised in practice
- Net message: you can’t reliably predict total cost from advertised per-token rates alone.

Model variant choice matters less than host reliability
- The same GLM 5.3 Flash model is served by many different companies (23 shown as available at the time).
- Speed varies by ~28x between fastest and slowest host.
- Pricing also varies by host; some can be far higher/lower than expected.
OpenRouter “Exacto” routing mode
- Designed to maximize tool-calling accuracy by avoiding hosts with high tool call error rates.
- A “tool call” is when the agent asks to read files or run commands; if the host fails, the agent stops.
- Exacto also uses signals like cache hit rate and tool call error rate.
DeepSeek Harness limitation: frozen model list
- DeepSeek Harness’s built-in model catalog is not updated from the web.
- Example discrepancy:
  - frozen list shows 333 OpenRouter models
  - OpenRouter sells 431
  - → 98 models are not visible to the harness
- DeepSeek’s own catalog reportedly shows only a few options in the frozen list.
- Plugins (e.g., FlyKit) can work around this by fetching a live list on startup.

DeepSeek Harness status
- Described as pre-release/alpha: 0.1.3 alpha.1
- Authors warn that compatibility-breaking changes may occur.
Why plugins can’t fully solve core-breaking issues
- Plugins can swap hosts/models, but if the harness core changes its rules, custom loops can break.

This setup doesn’t eliminate the need for a paid coding tool.
Key tradeoff
- You save ~10x money via cheaper setups/hosts
- but accept about ~15% worse chance that the host finishes requests (availability drops)
When it’s worth it
- Fine for background/overnight tasks, where retries don’t cost you time.
When it’s not
- Not ideal for deadline-critical work, because failed requests can cost more than the money saved.

DeepSeek
- DeepSeek Harness author/maintainer
- DeepSeek’s program and documentation/model guide
OpenRouter
- pricing
- weighted averages
- routing modes like Exacto
- host availability/tool-call error metrics
- cache hit-rate data
Anthropic
- Claude Code pricing
- claim that the free tier doesn’t include Claude Code
GitHub
- plugin repositories
- stars/forks
- version/tags
- pre-release alpha warning
Z.ai
- standing list prices for GLM 5.3 variants
Claude Code users/community plugin (mentioned)
- a plugin that authenticates a paid Claude subscription into DeepSeek Harness