The list KTPL kept
The Kirnani Technologies guide first ran on September 2, 2026, and was refreshed on October 26. The September version leaned on Google Flow and Veo. The October version does not. In their place: Hugging Face for open models and datasets, Tabnine alongside GitHub Copilot and Cursor, Wispr Flow for dictation, and n8n or Zapier for wiring APIs together.
General assistants stay where they were – ChatGPT, Claude and Gemini for drafting, reasoning and multimodal queries. Perplexity and NotebookLM cover research.
The reasoning behind the list is the part worth repeating: choose tools that fit a workflow you already have, not the ones trending this week. The guide also tells readers to verify output before it ships, code especially. That advice has not aged.
Where I part company is pricing. KTPL tells readers to check vendor pages for return on investment. Fair enough, but the vendor pages have moved a long way since September, and the spread between them is now the biggest single factor in what an AI feature costs to run.
One workload, ten prices
Take a mid-weight developer workload: 2 million input tokens and 500,000 output tokens per month. That is a heavy month of chat, code review and agent runs. Here is the same workload priced on today's published API rates.
| Model | Input $/1M | Output $/1M | Monthly cost |
|---|---|---|---|
| GPT-6 Astra | 10.00 | 50.00 | $45.00 |
| Claude Sonnet 5.5 | 2.00 | 10.00 | $9.00 |
| Claude Haiku 5.5 | 0.10 | 0.50 | $0.45 |
| Gemini 4 Argon | 2.00 | 10.00 | $9.00 |
| Grok 4.7 | 2.00 | 6.00 | $7.00 |
| Mistral Large 4 | 0.50 | 1.50 | $1.75 |
| GLM-5.3 | 1.40 | 4.40 | $5.00 |
| GLM-5.3-Flash | 0.15 | 0.50 | $0.55 |
| DeepSeek-V4.1-Flash (off-peak) | 0.15 | 0.60 | $0.60 |
| DeepSeek-V4.1-Flash (peak) | 0.30 | 1.20 | $1.20 |
Vendor list prices as of 26 October 2026: GPT-6 Astra $10/$50 from the OpenAI API pricing page; Claude Sonnet 5.5 $2.00/$10.00 and Claude Haiku 5.5 $0.10/$0.50 from the Anthropic API pricing page; Gemini 4 Argon $2.00/$10.00 from the Google API pricing page; Grok 4.7 $2.00/$6.00 from the xAI API pricing page; Mistral Large 4 $0.50/$1.50 from the Mistral API pricing page; GLM-5.3 $1.40/$4.40 and GLM-5.3-Flash $0.15/$0.50 from Z.AI's own pricing page; DeepSeek-V4.1-Flash off-peak $0.15/$0.60 and peak $0.30/$1.20 from the DeepSeek API pricing page. The GLM-5.3-Flash row uses Z.AI's own list price rather than aggregator figures.
Two things jump out. The first is the gap between tiers inside a single vendor: Claude Haiku 5.5 costs a twentieth of Sonnet 5.5 for the same tokens. The second is the gap between vendors at the budget end. GLM-5.3-Flash and DeepSeek-V4.1-Flash's off-peak rate both list $0.15 per million input tokens, roughly 67× below GPT-6 Astra's $10 input rate; DeepSeek's peak rate halves that gap to about 33×.
Those are dollar figures. Convert at recent rates, where the dollar buys a little under a euro, and every number in the USD monthly cost column drops by roughly 8%. I keep the table in USD because that is how the vendors publish it and because an exchange rate quoted today is wrong next month.
Where caching changes the arithmetic
Agents resend the same context over and over. That is why Google's introductory pricing on Gemini 4 Argon matters more than the headline: cached inputs get a 95% discount, which puts a repeated 20,000-token system prompt at $0.10 per million tokens rather than $2.00. DeepSeek splits its rate by time of day, so batch work scheduled overnight pays half.
Neither trick helps if your workload is one-shot prompts. Both help a lot if you are running a pipeline. The KTPL guide recommends n8n and Zapier for exactly that kind of multi-step automation, and the automation is where token bills are usually born.
Coding tools are wrappers around these numbers
GitHub Copilot, Cursor and Tabnine all sell an editor experience. Underneath, each one routes requests to a model from a table like the one above, and the vendor chooses which model you get. A subscription at a flat monthly fee hides that choice. It does not remove it: if the vendor swaps to a cheaper model to protect margins, your code quality changes without a changelog entry you will notice.
Tabnine has long positioned itself on where the model runs. Cursor sells speed and context handling. Copilot sells distribution. Try two of them on the same repository for a week before deciding, and check whether the subscription page says which models are in the mix.
Research, and what you upload
Perplexity searches the live web. NotebookLM answers only from documents you give it, which makes it far better at summarising a technical standard and far worse at telling you what happened yesterday.
There is a European wrinkle. NotebookLM works because you upload your own material. Upload internal design documents or customer contracts and you have handed them to a US provider. Under GDPR that is a transfer question, and it is the sort of thing a data protection officer will ask about before anyone gets an account.
The same applies to self-hosted n8n versus Zapier. n8n running on your own server keeps the data in your rack. Zapier is a SaaS product and your payloads travel to it.
What EU rules now require
Since August 2, 2026, Article 50 of the AI Act applies directly. Synthetic audio, video, image and text have to carry machine-readable marking, and users have to be told when they are interacting with an AI system. The "Digital Omnibus on AI", adopted in mid-2026, pushed the Annex III high-risk obligations to December 2, 2027, after CEN/CENELEC standardisation slipped. That delay does not touch the transparency duties.
In practice: if you build a customer-facing assistant on any API in the table above, you are the deployer. The disclosure duty is yours, not OpenAI's or Anthropic's. Enforcement of general-purpose AI rules by the European AI Office is also live now, with systemic risk assessments and training-data copyright documentation required from providers.
Mistral Large 4 is the one flagship in that table from an EU company, at $0.50 input and $1.50 output per million tokens. That puts it between Haiku and Sonnet on price, with an EU vendor on the invoice.
Open weights on your own hardware
DeepSeek-V4.1-Flash-1017 ships under the MIT licence. GLM-5.3 weights came out on August 28, 2026. For anyone who wants the marginal cost to be electricity, that is the route.
The constraint is hardware. Our own AI Arena rig runs a single RTX 5060 Ti with 16 GB of VRAM, and that capacity decides which of these checkpoints you can hold in memory and which need offloading or quantisation. Local inference buys privacy and predictable cost. It does not buy frontier quality at that memory size.
Measure before you commit
KTPL's own advice is the right place to stop: record the time and effort a task took before adoption, then after. A cheaper token is not cheaper if the model needs three attempts and a human correction. A $45 model is not expensive if it removes a review cycle.
None of the vendors publish a guaranteed latency figure next to those prices, and time-to-first-token moves with load. The only comparison that means anything for your team is the one you run on your own prompts.
Can the open-weight models run on a single consumer GPU?
Partly. DeepSeek-V4.1-Flash and GLM-5.3 both have published weights, but memory decides what you can actually load. On a 16 GB card you should expect to run quantised versions of mid-size checkpoints, with larger ones spilling to system RAM and losing speed. Check the checkpoint size against your VRAM before downloading anything.
Do I need to label answers from my own chatbot under the AI Act?
If the system interacts with people in the EU, Article 50 has required an explicit notice that they are talking to an AI since August 2, 2026. Generated audio, image, video and text also need machine-readable marking. The postponed deadline covers Annex III high-risk conformity obligations, not transparency.
Are these tools available and payable in the EU?
All the APIs listed above are reachable from EU member states, and most vendors bill in euro with VAT handling for business accounts. The exception to check per vendor is data residency: EU region hosting is offered by some providers and not others, so confirm it before sending personal data through an endpoint.
