GPT-6 Astra scores 72.6% on OSWorld at 40 minutes a task. For API teams, handing it the OpenAPI spec is faster, cheaper, and verifiable. When to click, when to call, and how to set it up.

INEZA Felin-Michel

September 5, 2026

We held off writing about GPT-6 Astra until we could test it ourselves. Two days later: the best model our team has ever run, what broke, what it costs, and why we say AGI is here.

Ashley Innocent

September 5, 2026

Gemini 3.8 Flash vs 3.7 Flash: same price, speed, and context, but +3 on the AA index, +12 on tau3-Banking, and 30% more output tokens per task. Upgrade?

Migrate from Gemini 3.7 Flash to 3.8 Flash: 9 API changes with before/after JSON, the minimal thinking-level error, call_id rules, token budgets, and rollback.

Ashley Goolam

September 3, 2026

Gemini 3.8 Flash thinking levels explained: what low, medium, and high do, why minimal now errors, per-level cost and latency numbers, and how to set each one.

Ashley Goolam

September 3, 2026

Gemini 3.8 Flash vs Claude Fable 5.1 vs GPT-5.6 Sol: specs, the 59/57/59 independent index, Google's benchmark rows, the 13x price gap, and which API to pick.

INEZA Felin-Michel

September 3, 2026

Gemini 3.8 Flash function calling step by step: declare a tool, read the function_call step, return function_result with call_id and name, cap loops, test.

Ashley Goolam

September 3, 2026

Step-by-step Gemini 3.8 Flash API guide: get an AI Studio key, call the Interactions API and legacy generateContent, set thinking levels, and test in Apidog.

Call gemini-omni-1.1-flash through Google's Interactions API: get a key, make your first curl and Python request, handle 4MB URI delivery, and save the call as a test in Apidog.

INEZA Felin-Michel

September 3, 2026

Fix the Claude Fable 5.1 400 "bound to a different conversation": what preserved thinking checks, who is enforced, drop_block, the audit, and append-only patterns.

Ashley Goolam

September 2, 2026

Use Claude Fable 5.1 in Claude Code: version 2.1.250+, model selection, what Pro and Max pay, effort settings, behavior shifts, and recommended CLAUDE.md lines.

Claude Fable 5.1 has no free tier. The honest paths: API signup credit, Max 50% weekly allowance, Pro usage credits, Batch at half price, and $0.25 cache reads.

Claude Fable 5.1 vs Opus 5: 2x the price uncached but cheaper cache reads, benchmark gaps from 1.4 to 23.6 points, where each wins, and a per-route decision rule.

Claude Fable 5.1 vs Fable 5: same price, cache reads cut to $0.25, every benchmark delta, 3 breaking changes, lost Priority Tier, and a decision rule by workload.

INEZA Felin-Michel

September 2, 2026

Migrate to Claude Fable 5.1 from Fable 5 or Opus 5: forced tool_choice 400, one-way thinking blocks, the history-editing check, every fix, and a full checklist.

Ashley Goolam

September 2, 2026