The price of artificial intelligence is moving in two directions at once. In the same week, OpenAI slashed prices on two of its models, Anthropic quietly canceled a planned price hike, and DeepSeek — the Chinese lab that built its reputation on cheap AI — raised its API prices by as much as 1,100 percent. Add Z.ai’s newly released GLM-5.3, an open-weights model with cybersecurity skills that reportedly grew faster than its own developers expected, and it’s clear the AI industry is entering a new, more volatile phase of competition. Here’s what actually happened, and what it means if you build on or pay for these models.
OpenAI Cuts Prices, Citing “Efficiency Gains”
On July 30, 2026, OpenAI reduced pricing on two of its GPT-5.6 models, as reported by Forbes. GPT-5.6 Luna, the company’s lightweight option, dropped 80 percent. GPT-5.6 Terra, a mid-tier model, saw a 20 percent cut.
| Model | Input (per million tokens) | Output (per million tokens) | Change |
|---|---|---|---|
| GPT-5.6 Luna | $0.20 | $1.20 | −80% |
| GPT-5.6 Terra | $2.00 | $12.00 | −20% |
OpenAI attributed the reductions to “under-the-hood efficiency gains” in how it serves its models, saying the changes would help “customers get more from every dollar they invest in AI.” The company did not frame the move explicitly as a response to competition, but the timing lines up closely with intensifying pressure from lower-cost alternatives out of China.
Anthropic Quietly Freezes a Planned Increase
Anthropic made a smaller but telling move, reported by The Stack. The company had planned to raise prices on Claude Sonnet 5 starting in September. Instead, it canceled that increase and kept the original introductory rate in place.
| Pricing (per million tokens) | Original / Kept Rate | Planned Increase (Canceled) |
|---|---|---|
| Input | $2.00 | $3.00 |
| Output | $10.00 | $15.00 |
Anthropic has not publicly explained the decision, but the effect is the same as a price cut relative to what customers were expecting to pay. Combined with OpenAI’s reductions, it signals that both leading U.S. labs are choosing to protect pricing rather than pass rising compute costs on to developers, at least for now.
DeepSeek Raises Prices by Up to 1,100% — After Building a Reputation for Undercutting Everyone
The bigger surprise came from DeepSeek. On August 13, 2026, the company launched DeepSeek V4 Pro, a 1.6-trillion-parameter mixture-of-experts model with a 1-million-token context window and adjustable reasoning levels, as reported by Caixin Global. It launched at aggressively low pricing, consistent with DeepSeek’s long-standing reputation as the budget option among frontier models.
Four days later, on August 17, that reputation gets more complicated. DeepSeek is set to implement price increases of 50 percent to 1,100 percent across its API, depending on the model, token type, and time of day. The company is also introducing peak and off-peak pricing for the first time, acknowledging that “demand is not evenly distributed” across its infrastructure and that variable pricing could push developers to shift non-urgent workloads to cheaper, quieter hours.
| DeepSeek V4 Pro | Launch Price (Aug 13) | Change Effective Aug 17 |
|---|---|---|
| Input (per million tokens) | $0.435 | Up to +1,100%, varies by peak/off-peak |
| Output (per million tokens) | $0.87 | Up to +1,100%, varies by peak/off-peak |
The timing is notable. DeepSeek has reportedly been raising funding at a valuation near $74 billion ahead of a possible listing on a Chinese stock exchange. Scaling infrastructure to meet real-world demand is expensive, and DeepSeek’s pricing shift looks like an acknowledgment that free or near-free inference was never a sustainable long-term strategy — even for a company that built its brand on it.
Z.ai’s GLM-5.3: A Coding Model With an Unplanned Cybersecurity Edge
While the pricing story dominated headlines, another Chinese lab released a model worth watching closely. Z.ai launched GLM-5.3 on August 14, 2026, describing it as its strongest open-weights coding system to date, according to Unite.AI. The company says GLM-5.3 uses the same base model as its predecessor, GLM-5.2, with all capability gains coming from additional post-training rather than a larger base model.
On harder independent coding evaluations, GLM-5.3 still trails frontier models such as Claude Fable 5 and GPT-5.6 Sol. But on Z.ai’s own benchmarks, the jump in both coding and security scores is large:
| Benchmark | GLM-5.2 | GLM-5.3 |
|---|---|---|
| Terminal-Bench 3.0 | 4.6 | 28.3 |
| DeepSWE v1.1 | 46.2 | 66.9 |
| CyberGym | 77.2% | 84.5% |
| ExploitBench | 24.4% | 54.4% |
| ExploitGym tasks completed (2 hrs) | 29 | 105 |
Z.ai says its models have already been used to identify 2,436 vulnerabilities across 269 open-source projects, including 1,097 rated critical or high severity.
Z.ai also disclosed something unusual: the model developed offensive security reasoning beyond what the company expected during training, moving from spotting individual bugs to forming coherent, multi-step exploitation plans. That prompted Z.ai to delay the public release of GLM-5.3’s model weights by two weeks for additional safety evaluation, with full weights expected around the end of August 2026.
Why This Matters: A Market Still Figuring Out What AI Should Cost
Taken together, these moves describe an industry without a settled answer to a basic question: what does running a frontier AI model actually cost, and who should absorb that cost?
A few patterns stand out:
- U.S. labs are competing on price stability. OpenAI and Anthropic are choosing to hold or cut prices even as they scale, likely to keep developers from migrating to cheaper alternatives.
- Chinese labs are shifting from land-grab pricing to sustainable monetization. DeepSeek’s price hike suggests that aggressive undercutting worked to build market share, but wasn’t built to last at scale.
- Capability gains are increasingly coming from post-training, not bigger models. Z.ai’s approach with GLM-5.3 — reusing a base model and investing in post-training — is a cheaper way to improve performance, and it’s a technique likely to spread.
- Security implications are becoming a pricing and access issue, not just a technical one. Z.ai’s decision to delay open weights shows that capability jumps in coding models can carry dual-use risk that companies now have to manage deliberately.
For developers and businesses building on these models, the practical takeaway is that AI infrastructure costs are not stabilizing — they’re becoming more dynamic, with peak pricing, tiered access, and sudden shifts becoming normal. Locking into a single provider without monitoring pricing changes is a riskier bet than it was even a few months ago.
FAQ
Why did DeepSeek raise its prices so dramatically?
DeepSeek says the increase, which takes effect August 17, 2026, reflects uneven demand across its infrastructure. The company introduced peak and off-peak pricing to encourage developers to shift less urgent work to cheaper time windows, rather than raising baseline prices uniformly.
Did OpenAI and Anthropic cut prices because of Chinese competition?
Neither company has explicitly cited Chinese rivals as the reason. OpenAI attributed its cuts to internal efficiency gains, and Anthropic has not commented publicly on why it canceled its planned increase. However, the timing coincides with growing competitive pressure from lower-cost Chinese models.
What makes GLM-5.3 different from other coding models?
Z.ai says GLM-5.3 reuses the same base model as its predecessor and gets its performance gains entirely from post-training. It posted large improvements on coding benchmarks, but its most notable trait is an emergent cybersecurity capability that grew faster than the company anticipated, prompting a delayed release of its open weights.
Is DeepSeek V4 Pro still cheaper than U.S. models?
At launch, yes — DeepSeek V4 Pro's initial pricing undercut most frontier competitors. But with price increases of up to 1,100 percent on certain models and token types taking effect days after launch, the cost advantage will narrow significantly for some workloads, especially during peak demand periods.
Sources: Forbes, The Stack, Caixin Global, Unite.AI.
