Changelog

Everything that changed in Fidget, newest first. Built in a day, on November 15, 2026 — and still going.

v1.9.7 November 16, 2026

  • The backend was still telling the AI that ChatGPT runs GPT-5, so recommendations kept naming GPT-5 even after the picks switched. The system prompt now names the GPT-6.1 lineup (Luna, Terra, Sol, Astra), so ChatGPT recommendations name a GPT-6.1 model.

v1.9.6 November 16, 2026

  • ChatGPT picks switched to the GPT-6.1 lineup: GPT-6 Luna (Instant), GPT-6.1 Terra (Medium), GPT-6.1 Sol (High), GPT-6 Astra (Extra High). GPT-5 and GPT-5 mini are gone from OpenAI's pricing page, so they no longer appear as recommendations. Prices come from the live sync and stay current automatically.

v1.9.5 November 16, 2026

  • Live model prices: Fidget now fetches the current prices straight from the providers' own pricing pages (Gemini docs, Anthropic pricing, OpenAI platform docs) and uses them for cost estimates, so intro-price expiries and model changes are picked up automatically. The catalogue refreshes in the background whenever it is more than a day old; if a fetch fails, the built-in prices are used instead. The admin page shows the last sync, the live prices, and any new or retired models it spotted.

v1.9.4 November 16, 2026

  • Effort rules tightened: prompts asking for full, comprehensive, thorough or in-depth research, a multi-part answer, or a biography covering several aspects now consistently get High or Max effort instead of sometimes dropping to Standard.
  • New fast pre-filter: before the AI sees your prompt, a deterministic filter fixes common typos, expands 'u' to 'you', strips leading filler like 'can you'/'please', and tidies spacing — instantly, with no network call. The panel shows what it changed. It never guesses the model or effort; the AI still makes that call.

v1.9.3 November 16, 2026

  • Fixed the copy-paste loop: after copying the optimized prompt and pasting it back, the panel asked to optimize it again. The optimized text is now recognised and left alone — the panel only re-checks when you actually edit it.

v1.9.2 November 16, 2026

  • Fixed a stale-recommendation race: if you kept typing while a recommendation was still loading, the slow reply for the old prompt could arrive after the new one and overwrite it, leaving the panel stuck on the old answer. A reply is now ignored once a newer prompt's request has started.

v1.9.1 November 16, 2026

  • Fixed the panel not updating when you keep typing: the change detector only looked at the first 60 characters of the prompt, so appending text past 60 characters never triggered a re-check. The panel now re-analyzes whenever the full prompt text changes.
  • Fixed the panel re-optimizing the same prompt over and over: it was watching every DOM change on the page, including its own re-renders, so each result it showed triggered another analysis. Changes inside Fidget's own panel are now ignored.

v1.9.0 November 16, 2026

  • Cloud-only: removed the on-device AI (Gemini Nano) fallback entirely. Every recommendation now comes from the cloud backend, the most accurate source. If the backend is unreachable, the panel says so — with a Try again button — instead of falling back to a weaker local model.
  • The Mode selector is gone from the popup (there is only one mode now), and the consent prompt no longer offers a local-only option.

v1.8.6 November 15, 2026

  • FAQ polished: merged the two overlapping prompt-data questions into one clear “What happens to my prompts?” answer, and added “How accurate are the cost estimates?” and “Which browsers does it work in?” — modelled on Token Wise's plain-English data split and honest limitations.

v1.8.5 November 15, 2026

  • Removed the local-rules tier entirely. The rules-based guess was the least accurate recommendation Fidget gave, and it fired exactly when the real AI was unreachable — the worst moment to guess.
  • Recommendations now come from two sources only: the cloud backend (most accurate) and on-device AI (the offline fallback). If neither is available, the panel says so — with a Try again button — instead of showing a guess.
  • On-device AI now classifies the task itself, so it no longer depends on the old rules to fill in the task, score, and tips.
  • Added this changelog page, linked from the footer.

v1.8.4 November 15, 2026

  • Added the changelog page at /changelog, linked from the footer.

v1.8.3 November 15, 2026

  • Landing page hero mockup re-skinned to Claude / Sonnet 5.5 / Medium, with real figures, so it matches the demo text above it.
  • Removed the “November 2026 snapshot” line from the footer.

v1.8.2 November 15, 2026

  • Prompt cache: the same prompt, provider, and mode now returns the same recommendation every time. Gemini isn’t fully deterministic even at temperature 0, so the cache is what makes repeat prompts consistent — and it saves a Gemini call on repeats.
  • The cache key includes the app version, so a release refreshes it. Entries expire after 30 days.

v1.8.1 November 15, 2026

  • Gemini calls now use temperature 0, for more consistent answers.
  • The popup footer version is now read from the extension itself — it was stuck showing 1.6.0.
  • Usage counts: the admin page now shows recommendations served in the last 7 days, by day, provider, and model picked.

v1.8.0 November 15, 2026

  • Cloud-first: the backend (Gemini) is now the first recommendation tier, because accuracy is the value. On-device AI is the fallback when the cloud is unreachable; local rules are last.

v1.7.2 November 15, 2026

  • Provider-aware validation: for ChatGPT, the model and effort are semi-independent, so a valid on-device model pick is kept. For Claude and Gemini, the model is derived from the effort.

v1.7.1 November 15, 2026

  • Fixed on-device recommendations returning a mismatched model and effort pair. The model is now derived from the validated effort.

v1.7.0 November 15, 2026

  • On-device AI: the extension can now use Chrome’s built-in Gemini Nano to recommend a model and rewrite your prompt entirely on your device — no network, no cost.
  • Recommendations now try on-device first, then the cloud backend, then local rules.

v1.6.7 November 15, 2026

  • Chrome Web Store readiness: permissions narrowed to exactly the three AI sites plus Fidget’s own backend, in-product consent before the first cloud call, a local-only mode, and a published privacy policy.

v1.6.6 November 15, 2026

  • Mobile optimization: fluid type sizes, a wrapping navigation, and full-width buttons on small screens.

v1.6.5 November 15, 2026

  • New landing-page story: “Your $20 plan, used like a $100 one.”
  • Added a section for teams — empower, don’t limit.

v1.6.4 November 15, 2026

  • Fixed the “only shows local” bug: the backend now sends CORS headers, so the extension can reach it from ChatGPT, Claude, and Gemini.

v1.6.3 November 15, 2026

  • Cleaned up the popup (removed the backend URL field and diagnostics) and the panel footer.
  • Rate limit raised from 60 to 300 requests per hour, per IP.

v1.6.2 November 15, 2026

  • The FAQ no longer says prompts go to Gemini — it says they go to “Fidget’s AI”.

v1.6.1 November 15, 2026

  • Fixed shared-prompt collection failing with a 502 (Upstash’s pipeline endpoint returns a bare array, not an object).

v1.6.0 November 15, 2026

  • Free for everyone: removed the 5-recommendations-a-day cap and the bring-your-own-key option. The backend does the AI now — no key needed — and the extension is shareable as a zip.
  • Redesigned the extension panel and landing page for scannability.

v1.5.0 November 15, 2026

  • Engineering foundation: one shared model catalogue for the web app and the extension, automated tests in CI, security lockdown (removed the debug endpoint and debug logs), and Upstash-backed rate limiting.

v1.4.0 November 15, 2026

  • Cost estimates now count hidden reasoning tokens, so the price shown for high-effort answers is honest.

v1.3.1 November 15, 2026

  • Fixed “Improve this prompt” failing with an “unexpected format” error; the call now retries, and the output cap was raised.

v1.3.0 November 15, 2026

  • Current model lineup and prices (November 2026): Gemini 3.8 Flash and 4 Argon, Claude Haiku 5.5 / Sonnet 5.5 / Opus 5.5 / Fable 5.2, GPT-5 and GPT-5 mini.

v1.2.0 November 15, 2026

  • Recommendations now name the real model for each interface, and the cost is output-aware.

v1.1.0 November 15, 2026

  • Added “Improve this prompt” — a rewrite of your prompt, with tips.
  • Recommendations run once after you copy, instead of re-running.

v1.0.0 November 15, 2026

  • First release. Fidget reads the prompt you’re typing in ChatGPT, Claude, or Gemini and recommends the right model and effort level, with a cost estimate and a tighter rewrite. The recommendation comes from Fidget’s backend; your own Gemini key is supported as a fallback.