OpenAI

1 months agoNews Editor

OpenAI to reopen $200 Pro plan as revised usage math cuts API-equivalent allowance

OpenAI said it will reopen its $200-per-month Pro subscription to new users tomorrow, while changing how usage is calculated under the plan. Tibo, a member of the OpenAI team, said that if the new Pro allowance is translated directly through API pricing, the usable amount comes out to roughly half of the old version. He also said that does not necessarily mean subscribers will get half as much work done, arguing that model efficiency and capability have improved enough that users could still complete more work than they could a month ago. The company does not plan to bring back the old five-hour cap. Instead, Pro users will continue to allocate usage across the week, with OpenAI adjusting the overall weekly usage calculation rather than imposing tighter time-window limits. Tibo said OpenAI wants the plan’s value to reflect real work completed instead of a headline API dollar amount. He also pointed to lower model costs, including recently introduced GPT-6 Sol and GPT-6 Luna API pricing that was cut 50% versus GPT-5.6 series promotional pricing. OpenAI said those lower prices stem from caching and inference efficiency gains, not a short-term promotion. Tibo added that OpenAI will announce more subscription-only features tomorrow, including some that will not consume usage.

740

Anthropic

1 months agoNews Editor

Claude Sonnet 5.5 leak points to gray testing, with pricing as low as $2 per million input tokens

Signals tied to Anthropic’s Claude Sonnet 5.5 are surfacing across X, where several developers say the model is already in late-stage gray testing inside Claude Code and could be released as soon as Tuesday afternoon U.S. time. Screenshots and configuration traces cited in the original report show a dedicated identifier, "claude-sonnet-5-5," alongside other model names including "claude-opus-5-5," "claude-sonnet-5," and "claude-fable-5.1." Early demos and side-by-side tests shared by developers suggest Sonnet 5.5 outperforms GPT-6 Sol in front-end UI generation, animation work, and day-to-day coding, while narrowing the gap with GPT-6 Astra in some advanced coding and agent-style tasks. One set of leaked pricing details has drawn particular attention: $2 per million input tokens, $10 per million output tokens, and $0.20 per million cache reads. The report also ties the timing speculation to OpenAI’s Sept. 29 DevDay, noting that Anthropic previously said Sonnet 5.5 and Haiku 5.5 would arrive in the coming weeks after the Sept. 22 release of Opus 5.5. Anthropic has not publicly confirmed the launch date, final pricing, or the full capability profile of Sonnet 5.5.

900

OpenAI

1 months agoNews Editor

OpenAI Teases a Heavy Release Week as GPT-6 Sol Speculation Builds

OpenAI is hinting at a packed week of launches, triggering fresh speculation that GPT-6 Sol could be part of the lineup. Product lead Tibo Sottiaux said the number of releases this week would be at the level people might have expected for DevDay 2025, and CEO Sam Altman followed with his own preview, saying a major version release is coming this week, with more significant announcements planned for DevDay. The company has not confirmed GPT-6 Sol. Still, the rumor cycle picked up after a non-official account that regularly posts Codex updates said the model could arrive on Thursday. Separate posts from users also claimed that some GPT-5.6 Sol traffic had already been routed to GPT-6 Sol in limited testing. One user described Sol 6 as heavily focused on reinforcement learning and speed, while another said early feedback from affected users had been broadly positive. At the same time, ChatGPT announced that GPT-5.5 will stop being available in ChatGPT, ChatGPT Work and Codex starting Oct. 14, requiring users to move to newer models. That change, combined with OpenAI’s recent releases across image generation, agents, real-time voice, enterprise data analysis and financial services, has added to the view that the company is preparing a broader shift in both its model lineup and product stack.

1970

OpenAI

1 months agoNews Editor

GPT-6 Sol surfaces in internal testing claims as OpenAI details agent-driven research acceleration

Claims about another OpenAI model surfaced shortly after Astra’s release, with online posts saying GPT-6 Sol is already being tested internally. According to those posts, Sol is weaker than Astra in overall output quality but much faster in at least one side-by-side task, where it finished in about one-sixth of Astra’s time while producing a similar amount of output. The report also cited a second Sol demo that generated a pixel-art sandbox prototype called "The Realm of Aurellune" in a single pass under zero-shot, Max-reasoning settings, using 60,000 tokens over 15 minutes. On the same day, OpenAI published internal research productivity data showing how heavily its own teams now rely on agents. As of mid-August this year, each researcher had about 3.1 agent workdays running in parallel for every 8-hour human workday, while the median researcher was using more than $600 a day in agent inference resources at API-equivalent pricing. OpenAI said it has reached the milestone of an "automated AI research intern" and set a new target of achieving an "automated AI researcher" by March 2028. Separately, chief scientist Jakub Pachocki argued that more capable systems are becoming harder to monitor through chain-of-thought inspection alone, and said the industry may need to slow model scaling if shared safety standards are not in place.

1150

Bit.Fan