AI Token Coach News Desk

5 items

• Thu 10 Sept

  ↓ BURN DOWN

  Model EfficiencyConfidence high

  DeepSeek ships V4.1-Flash at $0.15 per million input tokens

  MIT-licensed open weights, one million token context, and off-peak input at $0.15 per million. Cached input drops to $0.003 per million.

  How it reads for 3 stacks

  • ↓ Pays per token

    Off-peak input at $0.15 per million and cached input at $0.003 per million. If you run repeated-prefix work at volume, this is the largest single cost lever available this week.

  • ↔ Personal subscription

    Nothing changes on a subscription plan. This matters only if you also run metered work or are considering moving off a plan.

  • ↔ Company capacity

    Open weights under MIT may matter to your security team more than the price does. Worth raising if self-hosting is already on the table.

  SourcesDataNorth (primary)

• Thu 10 Sept

  ↑ BURN UP

  Model EfficiencyConfidence high

  OpenAI prices GPT-6 Astra at $10 and $50 per million tokens

  GPT-6 Astra lands at $10 per million input tokens and $50 per million output, matching Claude Fable 5.1 base rates. Fast mode costs 2x standard. Subscription users get it inside existing allowances, off by default at launch.

  How it reads for 4 stacks

  • ↑ Pays per token

    At $10 and $50 per million, Astra is 2.5x the current promotional rate on GPT-5.6 Sol. A workload of 1M input and 200K output costs $20 on Astra against $8 on Sol. Moving a default to Astra without testing is a direct invoice rise.

  • ↔ Personal subscription

    Astra is included in existing subscription allowances, so there is no new charge. It is off by default at launch, so nothing changes until you turn it on.

  • ↔ Company capacity

    Enterprise admins must enable Astra, and it is off by default. If you want it, the action is a conversation with your admin, not a settings change.

  • ↓ Codex

    Astra adds note-preserving context that Codex can search across prior windows, which should reduce the rework caused by losing debugging history to compaction.

  SourcesRuntimeWireOpenAI (primary)

• Mon 7 Sept

  ↓ BURN DOWN

  Agent and Harness EfficiencyConfidence high

  Claude Code weekly limits settle 25% above the original baseline on 14 September

  The temporary 50% boost ends on 13 September. A permanent 25% uplift over the original limit replaces it. That is more than the old baseline and less than what you have had lately.

  How it reads for 5 stacks

  • ↑ Claude Code

    Against what you have had during the boost, expect roughly 17% less weekly headroom from 14 September. If you regularly hit the wall, you will hit it sooner.

  • ↓ Personal subscription

    Against the limit your plan originally promised, this is a permanent 25% rise. If you joined before the temporary boost, you end up ahead.

  • ↑ Company capacity

    Seat-based Enterprise is affected. Consumption-based Enterprise seats are not. Check which one you are on before planning around it.

  • ↔ Pays per token

    API pay-per-token pricing is explicitly unchanged. This is a plan limit story, not a pricing story.

  • ↔ Claude

    Claude.ai chat limits are unchanged. This applies only to Claude Code in the terminal, IDE extensions and the desktop app.

  SourcesEveryday AIUsing Claude (primary)

• Sat 5 Sept

  ↓ BURN DOWN

  Agent and Harness EfficiencyConfidence medium

  Codex CLI 0.153 adds three-tier context management and a token budget mode

  Three-tier context management, an experimental token-budget mode, and Guardian history fencing. The third Codex release in a fortnight to target context spend.

  How it reads for 2 stacks

  • ↓ Codex

    Subagent budget accounting makes nested workers count against the root goal, which is the usual reason an agent run costs several times the estimate. Set a root budget and watch one real task before trusting defaults.

  • ↓ Pays per token

    Token-budget mode is experimental. Treat its figures as directional until you have reconciled them against your own billing.

  SourcesCodex Knowledge Base (primary)Codex Knowledge Base (primary)

• Wed 2 Sept

  ↓ BURN DOWN

  Router LabConfidence low

  LiteLLM AutoRouter v2 claims 45% lower cost per solved task

  A classifier that estimates each tier's odds of success and routes to the cheapest tier clearing a probability threshold. Benchmark claims 27% more tasks solved at 45% lower cost per solve.

  How it reads for 2 stacks

  • ↓ Pays per token

    The claimed 45% lower cost per solve comes from a 21-task vendor benchmark. The direction is plausible, the number is not yours until you have run your own tasks through it.

  • ↔ Personal subscription

    Routing across tiers assumes metered access to several models. On a single subscription there is nothing to route between.

  SourcesLiteLLM (primary)