GPT Assistant
Short answer: GPT-5.6 Sol is the best
quality choice for difficult coding work, but it is not the best single choice for performance, price, and very heavy usage combined.
For a high-volume Hermes coding workflow, I would use a tiered setup rather than run Sol at maximum effort for everything:
- Default for normal implementation, refactors, test fixes, and code review: GPT-5.6 Terra — Medium.
- Hard bugs, unfamiliar large codebases, architectural changes, security-sensitive work, or repeated failing tests: GPT-5.6 Sol — High.
- Maximum/Ultra: reserve for the genuinely expensive jobs: deep multi-file debugging, major migrations, difficult agent tasks, or a final “make this production-ready” pass.
- Luna: useful for cheap repetitive tasks—summaries, simple edits, grep-style investigation, documentation cleanup, and lightweight test-output triage.
If you want one default setting only: choose
Terra + Medium for Hermes. It should give the best practical throughput-per-dollar. Switch to
Sol + High when Terra starts looping, misses the root cause, or you need stronger planning and verification.
Why not Sol for every task?
Sol is OpenAI’s strongest GPT-5.6 coding model and has the larger context allowance in ChatGPT Work. It is the right “senior engineer” model. But direct API pricing is roughly double Terra’s: Sol is $5 input / $30 output per million tokens, versus Terra at $2.50 / $15. With agentic coding, output and reasoning can compound quickly—especially when Hermes is reading files, running tests, retrying, and iterating.
Terra is much closer to Sol than its price suggests, so it is the better default for an operation already doing hundreds of millions of tokens.
Effort recommendation
- Medium: best default. Fast enough, usually reliable enough, and avoids wasting long reasoning on routine tickets.
- High: use when the task has multiple dependencies, test failures are non-obvious, or a wrong change would be costly.
- Max: use sparingly. It is for difficult reasoning, not “better autocomplete.” At your token volume, it should be an escalation setting.
- Ultra: only for tasks where parallel sub-agents materially help. It is not a daily-driver setting.
One important current detail: in ChatGPT Business, the ChatGPT picker’s Medium, High, and Extra High options for GPT-5.6 Sol currently have the same listed flexible-price rate per message. However, higher-effort work can still consume more underlying agent/Codex resources and exhaust included usage sooner in coding workflows. So Medium remains the sensible default even where the visible per-message rate is the same.
About “unlimited” Business usage
It is
not truly unlimited in the way most heavy Hermes users mean it.
ChatGPT Business includes virtually unlimited access to eligible base/Instant usage, subject to anti-abuse controls. Advanced models and Codex have included per-seat allowances; once a user reaches those, your workspace needs credits to continue using that advanced capability. For three Business seats, you also have three separate included-usage pools before any shared credit spending.
Standard Business seats are currently about $25/user/month on monthly billing or $20/user/month on annual billing in the US, before tax. So three seats are roughly $75/month monthly or $60/month annual—not counting any optional flexible-usage credits. A $100/month target is reasonable only if you set a controlled credit budget for the team.
For Hermes specifically
If Hermes is configured through
OpenAI Codex OAuth, it uses your ChatGPT subscription/Codex access rather than a direct OpenAI API key. If it is configured with an
OPENAI_API_KEY, it is API-billed separately. Verify which provider Hermes shows in `hermes model`; that distinction matters far more than the token number Hermes reports.
Practical team policy
- Set Terra/Medium as the default coding agent.
- Allow Sol/High only for escalation or final review.
- Use Luna for low-risk bulk work.
- Turn off automatic top-ups initially, or set a small hard monthly credit ceiling.
- Track which jobs burn the most: long test loops, repository indexing, repeated retries, and verbose output are usually the first places to optimize.
My pick for your setup: Terra + Medium by default; Sol + High on demand. That gives you close-to-frontier coding quality without turning every routine Hermes run into a premium reasoning job.
