September 2026 AI Value: Opus, MiMo, DeepSeek and Luna

Historical September guide: This page preserves September buying guidance and dated corrections. Use the October AI buying guide for the newer shortlist; the September 27 OpenCode Go allowance section remains linked evidence with its own checked date.

October 3 OpenAI correction: Pro $200 purchases reopened; eligible existing subscribers keep their previous allowance through October 29, then move to a lower allowance at the same price. The launch/offers owner has the detailed terms. Sol 6.1 economics and included-capacity trials have separate current checks. This page preserves September buying guidance; its other-provider shortlist has not been re-reviewed for October.

Useful AI work does not require a $100–$200 subscription. As of September 27, Claude Pro with Opus 5.5 is a credible $20 starting point for focused work. MiMo V2.6 Flash, DeepSeek V4.1 Flash and GPT-6 Luna are strong value candidates for a cheaper primary route or fallback. Buy more capacity when the work you finish justifies it.

This is the dated buying shortlist, updated September 27, 2026. Use the cost-saving playbook for the lasting method and monthly audit; use the launch-offer guide for grants, expiry and DevDay reports. The August guide and the September 14 comparison below remain historical snapshots.

September 27 release update

Opus 5.5 makes the lower-priced Claude seat worth reconsidering. MiMo V2.6 Flash, DeepSeek V4.1 Flash and GPT-6 Luna give us a useful small shortlist across providers. These are recommendations to evaluate on real tasks, not a measured ranking of accepted-result cost.

The current OpenCode Go table also lists a $60 monthly V4.1 Flash limit on its $10 plan. Our older September 20 expiry note is not sufficient buying guidance today: see the fresh allowance check. Start with one route that suits your workload; a second recurring seat needs its own reason.

Workload router

Preserve model, effort, provider, harness version and task identity when comparing runs. The model guides own benchmark evidence and its limits. A cheap request that creates another hour of review is not the cheaper result.

Claude Pro is a real Opus work option

Anthropic lists Pro at $20/month in the US, including Claude Code; local prices and taxes vary. Current Claude Code model settings default Pro to Opus 5.5, replacing the older Sonnet default in recent clients. Max is not required merely to start doing useful Opus work.

Our recommendation is to try Pro for a bounded bug fix, review, research brief or writing task before buying Max. Start Opus at medium effort, keep relevant context and verify the output. Pro is a viable working tier when its allowance fits your workload; this is not a promise of all-day Opus capacity. Five-hour and other usage limits still apply, and Pro does not include separately billed Console API use. Check the model picker and usage meter in your account.

The Opus review explains why it merits the trial. The web/app billing checklist helps compare your final renewal total. Existing-subscriber cloud credits and saved resets can help with eligible work, but are separate from the recurring plan value.

Three inexpensive models worth actual work

These are direct API token rates, checked September 27, in USD per million tokens. They do not price a whole task or convert a subscription into API credit.

Sources: Xiaomi pricing, DeepSeek pricing, Luna model card. Cache hits, cache writes, reasoning output, tools and processing tier change the bill. DeepSeek’s peak schedule excludes weekends and Chinese public holidays. Luna applies a full-request long-context multiplier above 272K input tokens. Follow the owners for those details instead of comparing only the cheapest cell.

Pick one representative task, set a retry limit, and record whether the result passed plus review minutes. Use the existing route first if it is already paid for and suitable. None of these candidates has a site-owned CAR result in this release.

OpenCode Go: the current $10 / $60 distinction

OpenCode Go’s current terms, checked September 27, list $10/month and $60 of monthly DeepSeek V4.1 Flash usage, capped at $12 per five hours and $30 per week. That is a strong coding-value candidate if those windows fit your workload. It is provider-metered access through Go, not $60 deposited into a direct DeepSeek API account.

The current table also lists monthly usage limits of $60 for MiMo V2.6 Flash and $15 for GPT-6 Luna. Those are allowances within Go, not additional subscription prices. Do not add model rows into a promised combined balance: additive cross-model entitlement was not established. Go targets coding-agent traffic and requires client/session identification. Optional Use balance overflow spends a separate Zen balance.

Today’s page gives no V4.1 expiry date. We therefore record the current listing, not an assumed extension or permanent promotion. Limits can change; check renewal, final checkout total and the usage dashboard. The older September 14 terms remain below. For setup, use the OpenCode guide.

Start the weekly clock with useful work

Weekly windows are not a bank of unused weeks. Do not plan on two idle weeks giving you 2× allowance to spend back to back. Returning to an unused allowance also does not by itself mean another refill is one day away. Treat the displayed window and balance as your available capacity, unless the provider explicitly grants rollover or a separate credit.

For a usage-triggered weekly window, the clock starts with your first request. Leave it idle and you delay that start; when you return, the new week still has to run. Starting a small useful task earlier can therefore bring the next refill earlier than waiting until the day you need sustained work. That is a scheduling benefit, not an extra allocation. Do not send filler just to consume capacity.

OpenAI documents a concrete example after a paid Codex reset: allowance is restored immediately, but the new weekly period begins with the first subsequent Work or Codex request and refills seven days later. Two weeks idle after applying it do not bank two allowances or preserve the old refill date. This example does not make buying a reset necessary. For routine use, check the actual account clock before and after your first real task; OpenAI’s usage guidance points to that meter, rather than a universal calendar reset day.

Banked Codex resets also move the weekly date; Claude’s eligible saved reset instead preserves its usual weekly schedule. Keep those products separate. The capacity guide owns Codex mechanics, and the launch-offer guide owns Claude’s reset. Neither idle time nor a reset announcement creates a stored API balance by itself.

September 14 subscription comparison (historical)

This table preserves what we checked September 14, including the then-stated September 20 boosts. Use the September 27 OpenCode section above for today’s Go listing. Command Code has not been revalidated in this update. Dollar allowances are provider-defined usage values, not cash or portable API credit. Historical Command Code prices excluded a processing fee whose amount was unresolved.

Sources: OpenAI Pro terms, Command Code prices, GOAT, Pro, shared-window accounting, V4.1 offer terms, and OpenCode Go. Checked September 14, 2026.

Do not add Command Code’s model allowances together. Its usage guide describes one shared plan: lower-allowance models consume proportionally more credits per dollar. During the boost, an all-V4.1 GOAT workload therefore has $12/$30/$60 of model usage across the three windows, not $14/$35/$70. Pro’s equivalent is $14/$35/$70. Mixing models spends fractions of the same budget.

OpenCode documents model limits and 20%/50%/100% windows. This review did not establish whether switching models supplies independent additive balances; do not sum its catalog into a promised monthly entitlement. Advertised request counts assume particular cache-heavy requests, not complete coding tasks.

Clients, overflow, and expiry

September 14 record: Command Code documented CLI/API access for GOAT/Pro, paid overflow from extra credits and a billing-cycle refresh. The reviewed pages did not settle account eligibility, concurrency guarantees, expiry timezone or the treatment of an in-progress billing cycle. The old boost is not a current recommendation. OpenCode’s fresh terms are above.

A promotion helps only while you can use it. Compare ordinary renewal, correction time and client behavior as well as the headline allowance.

Pro retention and our Claude correction

Historical September 27 check: the then-reviewed Pro FAQ paused new Pro $200 purchases and offered an eligible one-time return within 30 days. October 3 correction: current terms reopen purchases and replace that return rule. Use the access record and offers owner before changing a plan. Retaining a productive seat still needs a case based on the allowance you will receive after October 29; $200 is neither our default starting budget nor a universal capacity advantage.

Our earlier Claude wording was too broad. Anthropic’s Max documentation says 5x/20x describes session capacity versus Pro, with separate weekly limits. It does not establish the same weekly multipliers or a universal “roughly double” weekly ratio between Max tiers. Any reported weekly ratio needs a dated, account-specific observation. See the Claude Max correction; this release makes no legal inference.

The author also reports verifying the Pro 20x purchase pause on September 14; this is account-level corroboration, separate from the provider notice. NVIDIA now lists Kimi K3, but that listing does not establish free or production access.

Measure useful work

Track API dollars, subscription consumption, output tokens, latency, and accepted-result cost separately. Count review time and retries, and divide attributable spend by accepted results. Avoid assigning a full seat price to one tiny experiment or treating unused allowance as savings. AIHackers has conducted no paid experiments for this release: CAR remains not-run.

Use the cost playbook’s monthly audit before renewal. This September 27 refresh checked the linked Pro access, model pricing, OpenCode allowance and reset sources; older benchmark and Command Code observations retain their dates. Recheck provider terms when acting and after a material change. Source archives remain archive-pending; no paid experiment or automatic source monitor is implied.