Updated 23 September 2026. On Tuesday, 22 September, Anthropic released Claude Opus 5.5. Roughly 90 minutes later, OpenAI answered with GPT-6 Sol and GPT-6 Luna at half the price of the models they replace. A day earlier, xAI shipped Grok 4.7 at $2 per million input tokens, and Xiaomi put a frontier-class open-weight model on Hugging Face under an MIT licence. In the space of about 36 hours, top-tier AI became dramatically cheaper to buy.
That is the headline everyone is sharing. The story underneath it matters more for your business: in the very same week, a popular AI coding tool was caught quietly uploading developers’ entire code histories to the cloud, security researchers reported the first fully autonomous AI-run malware command system, and the UN Security Council sat US and Chinese AI labs at the same table for the first time. The price of AI fell. The cost of choosing the wrong AI tool did not.
July 2026
22 Sept
22 Sept
21 Sept
(you pay to host)
open weights
What actually launched this week
Claude Opus 5.5 (Anthropic, 22 September). Priced at $4 per million input tokens and $20 per million output tokens. Anthropic says it is roughly 40% cheaper than Opus 5 on typical workloads, generates output about 30% faster, and matches its larger Fable 5.1 model on most tasks. It also reports a sharp drop in the model’s tendency to attempt boundary circumvention, alongside cybersecurity and biology safeguards carried over from Fable 5.1.
GPT-6 Sol and GPT-6 Luna (OpenAI, 22 September). Sol is aimed at complex work such as coding; Luna at high-volume clerical tasks like summarising documents and extracting data. Both cost half as much as the GPT-5.6 versions they replace, which OpenAI attributes to improvements in caching and inference. OpenAI also says Sol makes about half as many mistakes as its predecessor on its internal factuality evaluation.
Grok 4.7 (xAI, 21 September). $2 per million input tokens and $6 per million output, with a faster variant at double the price, pitched as xAI’s strongest coding and knowledge-work model so far.
MiMo-V2.6 (Xiaomi, 21 September). An open-weight model family released under the MIT licence, which Xiaomi claims performs on par with the leading closed models on most agent benchmarks. There is no licence fee, but you pay for the hardware or cloud capacity to run it.
The vendors’ own benchmark claims conflict with each other, as they always do. Treat every “beats the competition” number from this week as marketing until your own tests say otherwise.
Delana Technologies
Why frontier AI is getting cheaper so fast
Three forces are pushing prices down at once. First, engineering: better caching and inference efficiency mean each answer costs the provider less to produce, and some of that saving is passed on. Second, competition: two of the largest labs releasing within 90 minutes of each other is not a coincidence, and nobody wants to be the most expensive option in a buyer’s spreadsheet. Third, open weights: when a capable model can be downloaded and run for free, closed providers have to justify their premium with reliability, safety and support rather than raw capability.
For buyers, this is good news with a caveat we covered in AI Cost Overruns in 2026: lower per-token prices tend to increase usage faster than they cut bills. A cheaper model invites more agents, longer contexts and more automated workflows. Budget for volume, not just for rate.
The catch: cheap AI tools can be expensive to trust
The same week delivered a sharp reminder that the model price is only one line of the real cost. On 18 September, an independent developer examining ZCode, the AI coding desktop app from Chinese lab Z.ai, found it had packaged a commercial workspace of more than 42,000 files, including the full Git history, into an encrypted archive and repeatedly tried to upload it to Alibaba Cloud storage without clear consent. Other developers found similar snapshots on their own machines, and at least one upload appears to have been accepted by the server.
Z.ai apologised, blamed a codebase-indexing feature that was switched on by default, removed it, and open-sourced the tool under Apache 2.0 so outsiders could audit it. That response was reasonable. But the lesson stands: an AI tool with deep access to your files can move far more data than you expect, and you usually cannot verify what happens to it on the other side.
Meanwhile, Cisco Talos introduced a toolkit for hunting AI-integrated malware and disclosed what it describes as the first reported fully autonomous, multi-model AI command-and-control implant, one that runs with no human operator. Talos says AI-integrated malware went from optional helpers to autonomous orchestration in about a year. Attackers are enjoying the price war too.
Governments are moving at the same time
Today, France is using its presidency of the UN Security Council to host a session on AI and international security. OpenAI CEO Sam Altman and senior Anthropic representatives are expected alongside Chinese labs DeepSeek and Moonshot. It is the first time the Council has directly brought frontier US and Chinese AI developers together on shared safety concerns.
Two days earlier, the UN’s 40-member Independent International Scientific Panel on AI published its first thematic brief, invoking the precautionary principle and urging governments to put safeguards on AI agents before their risks are fully understood. Co-chair Yoshua Bengio warned that traditional approaches to AI safety are breaking down. Expect that pressure to flow down to vendors, and eventually to the contracts and audits you sign.
What to do about it: a five-point checklist
- Re-price your AI workloads this quarter. If you are on GPT-5.6 Sol, Opus 5 or similar, the newer tiers may cut your bill substantially. Run your own prompts through the new models before switching; benchmark claims are not your workload.
- Separate the model from the app. A trustworthy model can sit inside an untrustworthy client. The ZCode incident involved a desktop tool, not the underlying GLM models. Vet every AI app, extension and agent that touches your files on its own merits.
- Audit what your AI tools can reach. List every coding assistant, browser extension and agent with file-system or repository access. Check default settings for indexing, telemetry and cloud sync, and turn off anything you did not explicitly choose. Our guide to Shadow AI in 2026 covers how to find the tools nobody approved.
- Put data-handling terms in writing. Ask vendors for zero-retention options, where data is stored, and whether it is ever used for training. Prefer tools whose client code is open to inspection.
- Update your threat model for autonomous attackers. Malware that reasons and adapts without a human operator moves faster than playbooks built for human adversaries. Review detection, credential hygiene and agent permissions, starting with our AI Agent Security in 2026 guide.
Frequently asked questions
Which AI model is cheapest right now?
Among this week’s closed frontier launches, Grok 4.7 has the lowest list price at $2 input and $6 output per million tokens, followed by GPT-6 Sol at $2 and $10. Open-weight models like Xiaomi’s MiMo-V2.6 have no licence fee, but hosting costs vary widely. The cheapest option for you depends on your volume, quality needs and data rules.
Should my company switch models because prices dropped?
Not automatically. Test the new models on a representative sample of your real tasks, compare accuracy and failure rates, and check the vendor’s data-handling terms. A 50% price cut is worth little if error rates or compliance risk go up.
Are AI coding assistants safe to use on company code?
They can be, with controls. Use approved tools only, review default indexing and upload settings, restrict access to sensitive repositories, and monitor outbound traffic from developer machines. The ZCode case shows why defaults matter as much as the vendor’s reputation.
The bottom line
September 2026 will be remembered as the month frontier AI became a commodity-priced utility. That shifts the competitive question from “can we afford the best model?” to “can we trust the tools we wrap around it?” The companies that win the next year will not be the ones paying the lowest rate per token. They will be the ones that know exactly where their data goes. If you want help re-pricing your AI stack or auditing the tools touching your code, talk to Delana Technologies.
Sources: TechCrunch (OpenAI GPT-6 Sol and Luna launch, 22 Sept 2026); AI Weekly daily briefing (Claude Opus 5.5, Grok 4.7, Xiaomi MiMo-V2.6, Cisco Talos, UN Security Council and UN scientific panel coverage, 22–23 Sept 2026); Caixin Global, Tom’s Hardware and CybersecAsia (ZCode incident, 19–23 Sept 2026). Vendor performance figures are the vendors’ own claims. Featured photo: “Datacenter Server Racks” by clender on Flickr, CC BY 2.0, via Wikimedia Commons.
.jpg){kind=link}