Start with the final outbound request. Use the instruction field and message positions supported by that exact endpoint, preserve tool-call pairs, and verify the full task after the role error clears.
A Claude 400 on Amazon Bedrock can be a stripped beta header, old Claude Code, full context, a bad model ID or an account setting. Fix it by the exact message.
As of September 26, 2026, GPT-6 Luna and DeepSeek Flash cost least per token. Among $2-input models, Grok 4.7 is cheapest; past 272K tokens, Sonnet 5 beats Sol.
Creating a Claude API key is free; calls spend prepaid credits. New users get a small test credit, Max, Ultra, and Team add monthly API credits, and Pro adds none.
The Claude API bills your own code per token; Claude Code runs on Pro, Max or an API key. Max's new monthly API credits don't cover interactive Claude Code.
As of September 26, 2026, GPT-6 Sol matches Sonnet 5 and Astra matches Fable 5.1 on list price. Prompts over 272K favor Claude; Luna costs a tenth of Haiku 4.5.
As of October 10, 2026, heavy interactive Claude Code use is usually cheaper on Pro or Max than on API billing. Max's monthly API credits cover only automation.
The Claude Batch API halves every token price for jobs that can wait up to 24 hours, but on a big shared prompt, cached standard calls can still cost less.
Cloud session credit claims closed October 7, 2026; claimed balances still pay for cloud runs. Usage credits bill API rates; Max/Ultra/Team API credits fund the API.
Once you approve ANTHROPIC_API_KEY, Claude Code bills that key instead of your Pro or Max plan, and claude -p always uses it. /status shows which one pays.
Claude Code cache TTL is route-specific. Check the active route, TTL, cache creation tokens, cache read tokens, invalidators, and billing proof before blaming a bug.
As of October 10, 2026, no Claude Code coupon exists. Pro annual saves $40, bundles cut overflow credits up to 30%, and guest passes give 7 free days of Pro.
Claude Code runs Opus 5.5, Sonnet 5.5 and Haiku 5.5 at medium effort by default. Check with /effort status; /effort high saves per model, max lasts one session.
A ban after paying for Claude Max usually existed before the charge. Appeal at claude.ai/restricted while logged in and request any refund separately.
Pro is $20/month; Max (formerly Max 5x) is $100 and Ultra (formerly Max 20x) $200. Upgrade once limits, Fable or API spend cost more than the $80 or $180 gap.
"Rate limit reached" in Claude Code is usually a session, weekly, or Opus/Sonnet limit. Match the line, then resume in the same session without resending.
Run /usage for Claude Code's 5-hour session and weekly limits and their reset times. /status only shows account and model. Audit fast drain before upgrading.
Claude Pro gives more usage per session than Free, but as of October 8, 2026, Anthropic states no multiple. Test three real tasks before paying $20.
Claude extra usage bills at API rates: a 200K-token Claude Code turn costs $0.60 on Sonnet 5.5 and $3 on Fable 5.1. Over $80 a month on Pro, Max 5x usually wins.
Claude Fable 5.1 costs $10 per million input and $50 per million output tokens. Only cache reads got cheaper, $1 to $0.25, so savings track your cache use.
Opus 5 costs half as much for uncached input and output. Fable 5.1 costs half as much for cache reads. Your workload and acceptance rate decide which total is lower.
Claude Fable 5.2 outscores 5.1 at every effort level. On Artificial Analysis's long tasks it costs 32% less per task at medium, but more at high and above.
Opus 5.5 is the sensible starting point for most work. Fable 5.1 earns its premium when it solves problems that Opus still misses at higher effort—not simply because the job is long.
Fable 5.2 at medium nearly matches GPT-6 Astra at max for 39% less per task. On small tasks both passed, but Fable at high cost 1.74x what Astra did.