Which Cheap Frontier Model Deserves Your Money

The second week of July 2026 broke something in the AI market, and I mean that in a good way. OpenAI took GPT-5.6 public on July 9 in three tiers, and the cheapest one, Luna, runs $1 per million input tokens and $6 per million output. A day earlier, SpaceXAI, the merged SpaceX and xAI operation, shipped its first-ever model, Grok 4.5, at $2 in and $6 out. GLM-5.2 was already sitting there as the open-weight bargain of the season. Three genuinely capable models, all priced like commodities, all landing within about 48 hours of each other.

If your instinct was to open six benchmark tabs and start comparing win rates, close them.

The question that matters is smaller and far more useful: what is the cheapest model that clears the quality bar for the work you personally do every week? For most people that bar sits well below what launch-day marketing implies. A model that scores three points lower on some reasoning benchmark will summarize your meeting notes identically, and the money you save by admitting that is real money.

My position, stated plainly: most individuals and small teams should route the bulk of their daily work through something in the $1 to $2.50 input range, keep one stronger model on standby for genuinely hard problems, and then stop shopping until fall.

What each of these models is actually good at

OpenAI’s lineup is the easiest to reason about because the tiers do the sorting for you. Sol, at $5 input and $30 output per million tokens, is the flagship for problems where being wrong is expensive. Terra, at $2.50 and $15, is the sensible middle. Luna, at $1 and $6, is the workhorse. I walked through the full breakdown in our guide to the GPT-5.6 Sol, Terra, and Luna pricing tiers, but the short version is that Luna is startlingly good at the unglamorous 80 percent of knowledge work: drafting emails, summarizing documents, classifying support tickets, cleaning up transcripts, producing first passes of almost anything.

Terra is the sleeper of the three. It handles multi-step reasoning and mid-difficulty coding well enough that I have mostly stopped reaching for Sol unless something has already failed twice.

Grok 4.5 is a more interesting bet. At $2 in and $6 out, SpaceXAI is clearly pricing to buy market share, which works in your favor for now. It is also the company’s very first model, and first models from new labs tend to arrive with growing pains: thinner documentation, fewer integrations, tooling that assumes you enjoy debugging. Test it on a side project. Enjoy the subsidized pricing. Just do not move your whole workflow onto it in week one.

GLM-5.2 is the one I would tell any technical team to study hardest. Its backers pitch it as beating GPT-5.5 at roughly one-sixth of the cost, and while vendor claims always deserve suspicion, our GLM-5.2 deep dive found the value case holds up for a wide range of everyday tasks. Because the weights are open, it is also the only model on this list you can run on hardware you control.

And Google? Gemini 3.5 Pro remains stuck in limited enterprise preview, with wide release delayed. The delay itself carries a lesson, which we unpacked in our piece on what the Gemini 3.5 Pro slip signals: never plan your stack around a model that has not shipped. Buy what exists today.

So how do you match these to actual jobs? Think in terms of volume and difficulty.

• High volume, low difficulty: drafting, summarizing, classification, data extraction, reformatting. Use the cheapest thing that works. Luna, Grok 4.5, or GLM-5.2 all clear this bar, and a wrong answer costs you thirty seconds of review.
• Low volume, high difficulty: gnarly code, contract review prep, multi-step analysis, anything where a bad answer eats an afternoon. This is where Terra earns its price, with Sol as the escalation path.
• Long-form generation: reports, articles, documentation. Here the output price, specifically, decides your bill.

That last point deserves its own paragraph, because almost everyone reads the input price and stops. If your usage is chatty or long-form, meaning the model writes far more than you feed it, output tokens dominate what you pay. Luna and Grok 4.5 both charge $6 per million output tokens; Sol charges $30. Ask each to write the same fifty long reports and the flagship costs five times as much, and for straightforward prose you will struggle to tell the results apart. Flip the shape around for summarization pipelines, where you pour mountains of text in and get short answers back, and the input price becomes the number that matters. Figure out which shape your work is before you compare anything else.

When self-hosting beats paying per token

The Copilot news dragged this question into the mainstream. GitHub recently moved Copilot onto an open-weight model, which means teams can now self-host their coding assistant rather than renting it forever. That is a real shift, and it tempts people into assuming self-hosting is the frugal default.

Usually it is the opposite.

Self-hosting an open model like GLM-5.2 saves money under three conditions, and you want all three before you commit: your volume is large and steady enough to keep a GPU busy most of the day, you have someone on the team who genuinely enjoys running infrastructure, and you handle data that cannot leave the building. Miss any one of those and the per-token API wins, because the cloud provider spreads hardware costs across thousands of customers while your idle GPU sits there depreciating. An idle GPU is a subscription to nothing.

One more caution, and it is the one I feel strongest about. Every model switch carries a hidden cost. Your prompts stop working quite right, your intuition for the model’s quirks resets to zero, and you spend a week relearning where the failure modes hide. Jump tools every launch cycle and you pay that tax monthly while never building the fluency that makes these tools genuinely fast. The people getting the most out of AI right now are running boring setups they know deeply.

So here is what I would do this week, with real money on the table.

If you mostly write, summarize, and manage communication, make GPT-5.6 Luna your default and escalate to Terra when something is truly hard. That pairing covers 95 percent of what a solo operator or small office throws at a model, at close to the lowest price on the market. If you are a technical team doing serious coding or analysis, run Terra as the daily driver and evaluate GLM-5.2 seriously, self-hosted only if you pass all three conditions above, via API if you do not. Put Grok 4.5 in a sandbox and check back on it in a month, once the early rough edges get sanded down. Spend zero minutes on Gemini 3.5 Pro until it actually ships.

Then close the tab on launch coverage until October. The models will still be there, they will be cheaper than they are now, and you will have spent three months doing actual work instead of migrating between tools that are, for your purposes, nearly identical.