Model Guidelines
Last verified: 24 September 2026. Prices are USD. Check the model picker and usage dashboard before a long session; access varies by account, client and rollout. For model theory and hosting, see AI Models and Providers.
1. OpenAI / Codex Workload Guide
| Workload | Input | Expected output | Suggested model | Effort |
|---|---|---|---|---|
| Document reading | Documents / PDFs | Find, summarize, explain, compare | GPT-6 Sol | Medium |
| Document deep analysis | Documents / PDFs | Infer, investigate, resolve contradictions | GPT-6 Sol | High |
| Document + web | Documents + current sites | Verify, cross-reference, synthesize | GPT-6 Sol | Medium |
| Document + web deep analysis | Documents + web | Investigate complex/conflicting evidence | GPT-6 Sol | High |
| Document writing | Notes / sources / draft | Draft, rewrite, polish, structure | GPT-6 Sol | Medium |
| Analytical writing | Sources + ambiguous problem | Determine approach, analyze, write | GPT-6 Sol | High |
| Code analysis · ~100K LOC | Local repository | Locate code, architecture, flows | GPT-6 Sol | Medium |
| Code analysis · ~500K LOC | Local repository | Cross-module architecture/dependencies | GPT-6 Sol | Medium–High |
| Code analysis · ~1M+ LOC | Local repository | System-level and cross-module understanding | GPT-6 Sol | High |
| Development · ~100K LOC | Repository + requirement | Implement, test, validate | GPT-6 Sol | Medium–High |
| Development · ~500K LOC | Repository + requirement | Multi-module change, test, debug | GPT-6 Sol | High |
| Development · ~1M+ LOC | Repository + complex task | Investigate, design, implement, validate | GPT-6 Sol | High |
| Hard agentic development | Large repo + ambiguous problem | Autonomous investigation → design → implementation → validation | GPT-6 Astra | High; Max when justified |
Medium = execute a reasonably clear task. High = determine what the answer or solution should be, then execute it. Repository size increases navigation work but does not by itself justify the most expensive model. For code: repository → search/index → relevant files → dependencies → context → reason → act. OpenAI model guidance.
2. Claude Workload Guide
| Workload | Input | Expected output | Suggested model | Effort |
|---|---|---|---|---|
| Document reading | Documents / PDFs | Find, summarize, explain, compare | Claude Sonnet 5 | Medium |
| Document deep analysis | Documents / PDFs | Infer, investigate, resolve contradictions | Claude Sonnet 5 | High |
| Document + web | Documents + current sites | Verify, cross-reference, synthesize | Claude Sonnet 5 | Medium |
| Document + web deep analysis | Documents + web | Investigate complex/conflicting evidence | Claude Sonnet 5 | High |
| Document writing | Notes / sources / draft | Draft, rewrite, polish, structure | Claude Sonnet 5 | Medium |
| Analytical writing | Sources + ambiguous problem | Determine approach, analyze, write | Claude Sonnet 5 | High |
| Code analysis · ~100K LOC | Local repository | Locate code, architecture, flows | Claude Sonnet 5 | Medium |
| Code analysis · ~500K LOC | Local repository | Cross-module architecture/dependencies | Claude Sonnet 5 | Medium–High |
| Code analysis · ~1M+ LOC | Local repository | System-level and cross-module understanding | Claude Sonnet 5 | High |
| Development · ~100K LOC | Repository + requirement | Implement, test, validate | Claude Sonnet 5 | Medium–High |
| Development · ~500K LOC | Repository + requirement | Multi-module change, test, debug | Claude Sonnet 5 | High |
| Development · ~1M+ LOC | Repository + complex task | Investigate, design, implement, validate | Claude Opus 5.5 | Medium; High for difficult dependencies |
| Hard agentic development | Large repo + ambiguous problem | Autonomous investigation → design → implementation → validation | Claude Opus 5.5 | Medium → High |
| Frontier escalation | Long-horizon task that Opus 5.5 at higher effort could not solve | Reinvestigate, resolve, validate | Claude Fable 5.1 | High default; tune to task |
Claude’s adaptive thinking and model-specific effort controls are not identical to OpenAI’s. Sonnet 5, Opus 5.5 and Fable 5.1 support effort; Haiku 4.5 does not use the same control. Use the setting your Claude client exposes. These model choices are personal cost-conscious defaults, not a provider equivalence claim. Claude model lineup · Claude effort · Anthropic cost guidance.
Optional delegation prompt: “If this task has substantial independent tracks, delegate repository search, dependency mapping or test triage to bounded subagents. Use GPT-6 Luna for simple Codex scans, GPT-6 Sol Medium for harder Codex analysis, or Claude Sonnet 5 Medium for Claude workers. Have the lead agent integrate and verify their findings. Keep a single dependent change with one agent.” Subagents add token usage; use them when parallel work or isolated context helps. Codex subagents · Anthropic delegation guidance.
3. Models & Pricing
First-party Standard API text prices per 1M tokens, USD. Cached means a prompt-cache read. Output includes billed reasoning/thinking tokens. Relative cost compares uncached input and output rates with GPT-6 Sol = 1×; it does not compare quality, tokenizers or subscriptions. OpenAI API pricing · Claude API pricing.
| Provider | Model / API ID | Input | Cached input | Output | Reasoning / effort | Relative cost |
|---|---|---|---|---|---|---|
| OpenAI | GPT-6 Luna gpt-6-luna |
$0.10 | $0.01 | $0.50 | none–max; focused work | 0.05× |
| OpenAI | GPT-6 Sol gpt-6-sol |
$2 | $0.20 | $10 | none–max; general work | 1× |
| OpenAI | GPT-6 Astra gpt-6-astra |
$10 | $1 | $50 | low–max; hardest work | 5× |
| Claude | Haiku 4.5 claude-haiku-4-5 |
$1 | $0.10 | $5 | No effort control; fast work | 0.5× |
| Claude | Sonnet 5 claude-sonnet-5 |
$2 | $0.20 | $10 | Adaptive; general work | 1× |
| Claude | Opus 5.5 claude-opus-5-5 |
$4 | $0.20 | $20 | Adaptive; Medium default | 2×* |
| Claude | Fable 5.1 claude-fable-5-1 |
$10 | $0.25 | $50 | Adaptive; frontier work | 5×* |
*Cached-read ratios differ. Cache writes cost extra: GPT-6 and Claude 5-minute writes are 1.25× input; Claude 1-hour writes are 2× input. OpenAI GPT-6 prompts over 272K input tokens have higher rates for the whole request. Anthropic says its newer tokenizer can produce about 30% more tokens for the same text, depending on content. Tools, web search, regional processing and speed tiers may add cost. OpenAI pricing · Claude pricing.
4. Plans & Usage
| Provider | Plan | Price / month | Model access | Usage allowance | Coding / agent usage |
|---|---|---|---|---|---|
| OpenAI | Plus | $20 | GPT-6 Sol/Luna in Work/Codex; Astra access varies by rollout/client | Variable five-hour usage; Sol 15–150 estimated local messages | A few focused sessions/week; optional ChatGPT credits |
| OpenAI | Pro 5x | $100 | Plus models and Pro features | ~5× Plus; Sol 70–700 estimated local messages/five hours | Frequent sessions; optional credits |
| OpenAI | Pro 20x | $200 | Pro features; new sign-ups/upgrades paused as of 10 Sep 2026 | ~20× Plus; Sol 300–3,000 estimated local messages/five hours | Existing heavy-use subscribers |
| Claude | Pro | $20 | Current models; Fable 5.1 requires paid usage credits | Variable five-hour and weekly limits | Claude Code shares allowance with Claude |
| Claude | Max 5x | $100 | Current models; Fable included within a 50% weekly cap | 5× Pro per-session usage; weekly cap | Frequent Claude Code / research |
| Claude | Max 20x | $200 | Same model access as Max 5x | 20× Pro per-session usage; weekly cap | Daily intensive Claude Code |
Billing boundary: ChatGPT Work and Codex share included usage; extra ChatGPT credits have a separate token rate card. API-key use is separately billed in API dollars. Claude chat and Claude Code share plan limits; optional usage credits after the limit use standard API rates. An ANTHROPIC_API_KEY in Claude Code can switch the session to API billing. These systems have no trustworthy plan-to-token conversion. Check live usage before upgrading. OpenAI plans, limits and credit rates · OpenAI Pro $200 pause · Claude Pro · Claude Max · Fable plan rules · Claude Code billing.