i am building a map for understanding what’s being built across India. its live here india-infra-map.sunidhi-30.c… it covers 7,000+ infra projects and lets you - filter by state, category and reported status - see who is responsible and what it costs - find projects past their reported deadlines - track major cost and timeline changes - see where the evidence is missing let me know what you want to see next @India_Stack @IndianTechGuide

intelligence too cheap to meter 15 months later, Luna scores higher than GPT-4.1 even without reasoning, with 10× cheaper input and 6.7× cheaper output

gentle reminder: if you exhaust your regular Codex limit, GPT-5.3-Codex-Spark still has a separate allowance on Pro

GPT Luna leads the price-performance frontier. It currently offer's the most intelligence per dollar for high-volume agentic work and for well defined clear tasks Luna seems like an obvious default. Depending on reasoning effort, it costs roughly $0.01–$0.06 per task. and at max mode, Luna reaches an Intelligence Index score of ~51 for ~$0.06 per task. if you compare other models - - GLM-5.2 Max: ~4× more - Claude Opus 5 Low: ~6× more - Gemini 3.6 Flash: ~8× more Another surprising thing is that at roughly $0.02–$0.03 per task, Luna already reaches ~46–49. So paying twice as much for max effort adds only another two points.

cost curve changes are equally as important as intelligence curve changes for instance, GPT-5.4 xhigh and GPT-5.6 Luna max both score 51 on Artificial Analysis. 5.4 costs $2.50/$15 per million tokens. Luna costs $0.20/$1.20. its insane that the same benchmark score is now 92% cheaper, in under five months.

DeepSeek says V4 Flash kept its 284B-total, 13B-active architecture and was only re-post-trained. Terminal Bench: 61.8 --> 82.7 DeepSWE: 7.3 --> 54.4 At $0.28/M output, it's 89× cheaper than Opus 4.8. Different harness, so not a clean before/after. Still, wild price-performance.

DeepSeek V4 Flash is getting surprisingly close to Opus 4.8 on coding evals: DeepSWE: 54.4. vs. 58.0 Terminal Bench 2.1: 82.7 vs 85.0 Price: $0.14/M input, $0.28/M output. note that - these are DeepSeek-reported harness results, but the price-performance is absurd.

After using ChatGPT Work for a week, I think we're watching knowledge work get abstracted in real time. Codex began this abstraction for software engineering. ChatGPT Work is trying to extend it to knowledge work itself. Working in large enterprises meant constantly switching between projects, apps, files, and internal tools. Gathering context, researching things, writing code, and sometimes just trying to understand which tools even existed inside the company. Half the work was figuring out the right tool for the right job. ChatGPT Work starts collapsing all of that. You can give it an entire project, connect things like Slack, Notion, and Google Drive, and ask it to act almost like your chief of staff. It can gather context, keep track of what is happening, create things, and continue working until it actually needs a decision from you. Once you have a repeatable workflow, you can schedule it instead of doing the same work manually every day. It can continue in the cloud, and you can pick it up from another device. Our job is moving one layer higher. Instead of operating every tool ourselves and translating one goal into 20 different steps, we increasingly just explain what needs to get done and review the work. Even the amount of prompt engineering feels noticeably lower going from GPT-5.1 to GPT-5.6. Less time explaining every step and more time describing the intent. It is kind of crazy to watch this happen in real time.

i still don’t understand why enterprises haven’t adopted ChatGPT Sites more aggressively the stark reality is that most companies spend recklessly on SaaS. some have thousands of subscriptions and spend millions, sometimes even close to a billion dollars, on software every year with ChatGPT Sites, you can literally prompt and build custom tools for your exact use case instead of buying yet another SaaS subscription in many cases, the money you save on SaaS can pay for the AI itself. just redirect some of that budget into AI access and keep building the tools you actually need

ChatGPT Work is how most users will have a first taste of AGI

most people won't discover AGI via a benchmark they'll discover it the first time they give ChatGPT Work a messy project, close their phone, and come back to a finished spreadsheet, deck, report or website.

built with ChatGPT Sites + GPT-5.6 Sol its a map where you can explore airports, ports, railways, highways and other critical projects across the country, then dig into their costs, timelines, status, responsible organizations and sources india-infra-map.sunidhi-30.c…

We've had 10 Codex resets in the past 21 days, so I'm not sure how sustainable this cadence is. Either the resets become less frequent and we're left token-starved or OpenAI continues making inference dramatically more efficient, which I think it will, considering how much efficiency has improved over the past four years. To be honest, there is barely a minute when I don't have Codex running and I still exhaust my limits in less than two days. I'll probably have to move from xhigh to high or even medium and low, use standard mode always or rely more on smaller models like Terra and Luna. But asking users to manually choose between xhigh, high, medium, Terra, and Luna for every request is inefficient and creates a poor user experience. I'd love intelligent routing that automatically chooses the right model and reasoning level for each task. It would save tokens, inference costs, and compute. I also wish OpenAI offered a bigger subscription, maybe a $500 plan. It would already be worth it for me. If OpenAI can make GPT-5.6 on xhigh 100 times cheaper over the next four or five months of engineering, we'll see exponential gains. The capability already feels solved. What remains is making inference efficient enough and giving power users the option to pay for more of it.

its too inefficient for the user to choose between xhigh, high, medium, Sol, and Terra for every task. Just route each request to the most efficient model automatically and let them keep building.

that makes it 10 in past 21 days - thanks Tibo chatgpt-codex-reset-counter.…

- Delhi–Amritsar–Katra (DAK) has 11 tracked package entries - Original cost is ₹18,529.75 cr, revised to ₹19,596.84 cr (+5.8%). - Deadlines are staggered, some shifted toward 2028, and one package is still in Balance-for-Award. you can track it here for more details - india-infra-map.sunidhi-30.c…

India has 24 airports under construction - you can now track its cost, timeline, contractors, extensions and more here - india-infra-map.sunidhi-30.c…

update: we now have 10k+ records and you can now filter by projects that - past reported deadline - cost atleast 2x now - 10+ year extension india-infra-map.sunidhi-30.c… explore and let me know what else you want to see

update: we now have 10k+ records and you can now filter by projects that - past reported deadline - cost atleast 2x now - 10+ year extension india-infra-map.sunidhi-30.c… explore and let me know what else you want to see