Australian engineers shift from Fable 5 to GPT-5.6

Antiburn says engineering teams in Australia are shifting from Anthropic's Fable 5 to OpenAI's GPT-5.6, citing usage and cost data from about 1,000 engineers.

The data suggests a sharp change after Fable 5's relaunch. Antiburn, which manages AI coding tools for engineering teams, can track usage and spending directly across those systems. It says some organisations that had relied exclusively on Anthropic models have also cut back their use of Fable 5.

Five-hour usage limits are now being hit by 20% or more of Fable users, up from 3%, Antiburn says. It also says Fable 5 costs four times more in tokens than Anthropic's Sonnet 5 for the same work, using about 2 million tokens on average to complete tasks that Sonnet 5 handles in roughly 500,000.

The shift reflects a wider concern in the AI market over whether more expensive frontier models deliver enough additional value to justify the cost. Developers and companies have been weighing premium models against cheaper alternatives, especially as rivals release models with similar performance on common coding and business tasks.

Dave Slutzkin, Chief Executive Officer and Co-Founder of Antiburn, said the issue is not simply model quality, but the economics of using it at scale.

"Anthropic's Fable 5 launch was a case of the company overpromising and underdelivering. The original version of Fable 5, which many users loved, was very strong, but it was only live for a couple of days. The version they shipped is disappointing by comparison and too expensive for many. Following the relaunch, our data shows users rapidly jumping to GPT-5.6 in the wake of the post-nerf sadness, even in organisations that were previously Anthropic-only.

"The real problem for Anthropic is token burn. Overall, it's a strong model, but it chews through subscriptions and token bills far too fast. We're seeing 20% or more of users hit five-hour limits with Fable, up from 3%. Most of those people have now throttled back to Opus 4.8, so they're getting no benefit from the new model at all.

"This has been a strategic misstep. OpenAI's 5.6 model turned out to be at least as good and much cheaper, so Anthropic has now been forced to include Fable in subscriptions, which it never intended. Fable was supposed to be its pre-IPO cash cow and instead it's draining revenue - very expensive to serve, with no marginal revenue associated with it," Slutzkin said.

Engineer views

Other Australian technology executives described a narrower set of tasks where Fable 5 stands out, while saying most routine work can be done more cheaply elsewhere.

Jordi Hermoso, Co-Founder and Chief Technical Officer at Doccy & Medlo, said his team found the model useful mainly for specialised technical analysis and architectural work, rather than day-to-day engineering.

"Fable 5 is exceptional at a very specific category of work. We've used it to review academic cybersecurity research and produce surgical recommendations for our own system. It's good at traversing a sprawling codebase and proposing architectural changes that genuinely simplify it.

"But it's not my daily driver, and it's not most of my team's either. I surveyed our engineering team and 80% hadn't found a compelling use case for Fable 5 and had stopped using it. The 20% who love it only pull it out for rare, highly specialised jobs.

"For a few specialised jobs, Fable 5 has a real lead. The problem is how few of those jobs there are. It's a scalpel, and most work doesn't need a scalpel. Fable has a real lead on a narrow set of tasks, but I don't expect it to stay exclusive. Open-source distillation plus the race toward cheaper and faster models will keep narrowing the gap month by month.

"Anthropic has mastered hype-driven product launches. The marketing runs well ahead of what the product actually delivers. Trust is complicated when every lab claims the moral high ground. Anthropic's communications around usage still land too close to the Ministry of Plenty: the credits drop, the previous higher number is erased, and the reduced ration is sold as good news.

"The moment the US government pulled Fable access after the three-day window, it felt like superintelligence had been switched off. Losing it, even briefly, revealed how quickly teams begin treating frontier AI as a permanent given. That dependence is worth facing honestly, especially now the wider market is catching up and the capability itself is probably never going fully dark again.

"We're not team Anthropic or team OpenAI. We're deliberately model-agnostic and use the sharpest tool for each task rather than betting the company on one lab's roadmap," Hermoso said.

His comments suggest that, for some engineering teams, the debate is shifting from which model is best in absolute terms to which offers the best fit for a specific job. That can favour a mix of premium and lower-cost models rather than a single provider.

Joe Gibbs, Lead Technology Engineer at Pressto AI, said businesses are likely to focus more closely on return on investment as model pricing becomes harder to absorb.

"I use AI tools every day, and I've found that Claude Fable 5 is great, but only worth using for the hardest tasks. It's definitely better than GPT-5.6 Sol, but that's still a close alternative.

"Anthropic had been extending the period that Fable could be used on a subscription because it knew that if it charged API prices, it would lose customers. To test it out, I tried using Fable with API billing, and it cost $50 for two questions. Meanwhile, GPT-5.6 lets me do eight hours a day of work for $155 per month.

"Fable 5 recently became a permanent part of Claude's higher-tier subscriptions, but the $20 Pro subscription still needs API billing. I think it's good to use if you have some tough questions, as long as it's used alongside cheaper models.

"While I do use the big models for personal use, businesses should focus on training small models to serve customers if they want to avoid massive bill shock. Small models might not be perfect all-rounders, but they're much easier to host and can give users what they need faster, without the worry of your data being used for training," Gibbs said.

The comments underline a growing divide in the AI software market between flagship models that can handle specialised, complex work and lower-cost systems that are increasingly good enough for mainstream use. For buyers, that makes pricing discipline and token efficiency as important as raw model performance.

Antiburn's figures indicate that, at least among the teams it monitors, those trade-offs are already changing purchasing behaviour. Five-hour usage limits are being hit by 20% or more of Fable users, up from 3%.