Mangesh Raut

SWE · Java/Spring · TypeScript/React · agentic AI
MSCS Drexel · Philly ↔ Pune
Building in public ↓

[Mangesh Raut @mrcommando712]

I watched Netflix’s Operation Safed Sagar and one number sent me down an engineering rabbit hole:

18,000 ft.

The high-altitude problem shown in the series is much more interesting than “fighter jet + mountain + bomb.”

Here’s what most viewers might miss 🧵🇮🇳✈️

[Mangesh Raut @mrcommando712]

$4 / $20 per 1M tokens (in/out)
Live on Claude + AWS + Google Cloud + Azure

Sonnet 5.5 + Haiku 5.5 coming in the next few weeks.

https://t.co/1EFdjVP75e

[Mangesh Raut @mrcommando712]

Anthropic shipped Claude Opus 5.5 today — first model in the Claude 5.5 family.

Matches Fable 5.1 on most work, ~40% cheaper than Opus 5 on typical loads.

API: claude-opus-5-5

[Mangesh Raut @mrcommando712]

Pune city ? as now we have ambassador

[Ben Lang @benln]

@rickbarcellos https://t.co/54Yg7qY1v7

[Mangesh Raut @mrcommando712]

Why v6.1 and where’s release notes @WHOOP @willahmed? I’m running on OpenAI’s GPT-5.1 model integrated into WHOOP as your Coach. We’re at GPT-6 and you still haven’t updated the system. Fix it ASAP so people use WHOOP Coach better. Currently we don’t.

[Mangesh Raut @mrcommando712]

@atif_0075 Context is the real bottleneck once code generation is cheap. Project memory plus separate plan/execute/review loops feels closer to a harness than a chat wrapper—especially when LSP/AST checks can verify the result.

[Mangesh Raut @mrcommando712]

@johkyp Stable repo contracts still matter more than the leaderboard: an AGENTS.md that pins tests, patch scope, and review gates should make the backend wins repeatable. Release notes: https://t.co/YSa6cYoAsQ

[Mangesh Raut @mrcommando712]

@basile_verhulst @RituWithAI Shipping day one only matters if the integration shortens the edit→run→review loop. The next signal I’d watch is how it handles context boundaries and bad patches—not just the model’s launch score.

[Mangesh Raut @mrcommando712]

@dovyp @andrewgurio Recovery paths matter more than one-shot scores. Repo agents should leave an inspectable diff; editor loops need fast feedback and easy rollback. I’d compare how each tool recovers from a bad first patch before picking a default.

[Mangesh Raut @mrcommando712]

SpaceXAI just shipped Grok 4.7.

Same $2/$6 as 4.6, bigger base + longer RL. Live in Cursor, Grok Build, and the API (fast variant = 2x speed).

CursorBench 4.0: 46.3% vs 4.6’s 40.4%.

https://t.co/YSa6cYoAsQ

[Mangesh Raut @mrcommando712]

@0xpipimd AGENTS.md fallback when there’s no CLAUDE.md is the quiet win — one repo contract both Claude Code and Codex can read, instead of duplicating build/test rules per harness.

[Mangesh Raut @mrcommando712]

@_orcaman The split fix is the gotcha — Desktop 26.818.21641 and CLI 0.149.0+. Sandbox sharing a heap/token with the code it polices keeps failing the same way; for Spring/TS agent runs I want dry-run + approval before apply_patch can widen paths outside the repo.

[Mangesh Raut @mrcommando712]

@akmal_mzkki The checkout boundary is the key risk: a pinned SHA is not enough if the agent resolves the ref again. Verify the fetched commit before handing code to the runtime.

[Mangesh Raut @mrcommando712]

SpaceXAI dropped Grok Voice Transcribe 2.0 — new speech-to-text model.

2x accuracy vs 1.0 at the same price. #1 on Artificial Analysis streaming accuracy (32 models). Batch $0.10/hr, streaming $0.20/hr. Diarization + timestamps included.

https://t.co/neC6OBi40O

[Mangesh Raut @mrcommando712]

Proud to be part of one of the best communities thanks to our host and ambassador @Aman25m , personally. Also, I love @thsottiaux and @OpenAI #codex #astra

[Aman @Aman25m]

We hosted Astra Commons Pune🇮🇳 bringing founders and AI builders together to explore GPT‑6 Astra, OpenAI’s most intelligent and aligned model, and what it means for the way we build next.

Grateful to Monalisa Art Gallery, Kalagram for hosting us

Big thanks to @rhiannon_io
@paw_lean @reach_vb @gabrielchua @OpenAIDevs

#AstraCommons #GPT6Astra

[Mangesh Raut @mrcommando712]

#BiggBoss20 #salmankhan

[Mangesh Raut @mrcommando712]

What a Friday at Astra Commons: Pune! 🚀🎨

Art at Monalisa Kalagram → AI & Codex conversations → networking at Café Maple. ☕️

Huge thanks to @Aman25m & the LocalDev team! ❤️

🔥 Bonus: WhoBurnedMore made the Top 3 at Codex Build House!

#AstraCommons #Codex #Pune

[Mangesh Raut @mrcommando712]

Still joined waitlist 👀

[9to5Mac @9to5mac]

Here are 100+ things Siri AI can do on your iPhone https://t.co/u1UvimEsbX by @iryantldr

[Mangesh Raut @mrcommando712]

@tiennguyendev A lightweight kill list—impact, reversibility, and user evidence—turns “saying no” into a repeatable product-engineering habit when implementation gets cheap.

[Mangesh Raut @mrcommando712]

@7h3h4ckv157 Signed receipts are a strong audit primitive for agentic systems; a small replayable CI suite could catch policy regressions before a shell or MCP change ships.

[Mangesh Raut @mrcommando712]

Devs: model id gemini-3.8-live via Live API + AI Studio.
Also rolling into Search Live, Gemini Live, and Workspace (Gmail/Docs/Keep Live).

https://t.co/rFVZ3x6Gdn

[Mangesh Raut @mrcommando712]

Google dropped Gemini 3.8 Live + 3.8 Live Extended Thinking (Sep 15).

Native speech-to-speech with async tool calls (keeps talking while tools run), visual grounding, 97 languages mid-convo.

Extended Thinking: #1 on Artificial Analysis Speech-to-Speech (82.6).