SShortSingh.
Back to feed

OpenAI's GPT-6.2 Astra Cuts Base Rates, but Hidden Reasoning Costs Test Developers

0
·2 views

OpenAI has updated its model lineup, lowering base rates for GPT-6.2 Astra by about 20 percent and adding a discount on cached input tokens. Developers weighing a switch report that hidden reasoning tokens can erase those savings, and one analysis works through the cache breakeven point using concrete per-million-token prices. Influencers and commenters on Hacker News say the model behaves erratically on routine tasks and trails competitor models on some benchmarks, with skepticism about chart presentation and release timing. Critics argue that benchmarks measure answer quality but ignore the token cost of producing it, which matters most in production workflows. The discussion concludes that developers should evaluate cost per completed task rather than headline benchmark scores.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

H200 GPU rental prices rise as memory shortages strain AI compute supply

Reported H200 GPU prices have climbed even though the chip is not new, and sources disagree on the exact size of the increase. The H200's 141GB of HBM3e memory, versus 80GB on the H100, makes it attractive for large-model and long-context workloads, while DRAM supply is expected to stay tight through 2027 and into 2028. The article argues the hike signals that AI compute is scarce, memory-constrained and volatile, ending the assumption of steadily falling prices. For steady workloads, the added monthly cost can erode project margins, so teams are advised to budget with a buffer, use only needed memory, cut idle hours and compare quotes often.

0
ProgrammingDEV Community ·

Free UK VAT calculator exposes a common freelancer error in removing VAT

A developer has built a free online VAT calculator aimed at UK freelancers who often miscalculate VAT. The article explains that removing VAT from a gross price by subtracting 20% is wrong, since the VAT is calculated on the net amount, not the gross. It notes that UK businesses must register for VAT once taxable turnover exceeds £90,000 in any rolling 12-month period, and that this threshold is measured on a gross basis. The calculator lets users add or remove VAT at 20%, 5% or 0% and shows the net, VAT and gross figures. The author warns that errors in either direction are costly, as undercharging means paying the tax from one's own pocket and overcharging effectively makes clients pay tax on tax.

0
ProgrammingDEV Community ·

Product Hunt launch drew 1 upvote and zero sales for a creator storefront

A creator storefront called AEFRI scheduled a Product Hunt launch but received only one point and no comments. Its checkout platform Payhip recorded 218 visits and 105 product views month to date, with no paid checkout starts and $0.00 in revenue. Most traffic was direct, with search referrals totaling only 18 visits from Google, Yahoo, Baidu, Yandex and Bing. The team attributes the outcome to having no existing audience, since Product Hunt rewards early engagement that the launch lacked. They have since added a launch coupon, a free-sample call to action, and a fourth free tool, and they are focusing on search-driven growth over the coming months.

0
ProgrammingDEV Community ·

Developer releases free dataset of recorded AI coding sessions, including failures

A developer is offering 66 recorded coding sessions, comprising 2,541 real tool calls across five open-source repositories, under an MIT license at no cost. The recordings capture full work sessions, including failed test runs and their fixes, rather than just final code diffs. The author argues that AI companies have profited from developers' code and that this data should be shared freely instead of sold. They report that training a small Qwen model on the data raised correct tool calls on their own test from 43% to 80%, while other benchmark scores worsened, and they have published these results openly. The dataset includes a verification script so users can confirm the files match what was published.