Workhorse shootout: Gemini 3.7 Flash, GPT-5.6 Luna, Sonnet 5

Google made gemini-3.7-flash generally available yesterday, 13 August 2026, calling it "our most intelligent workhorse model yet for coding and agents" and attaching an introductory price of $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026. That release lands in the middle of a remarkable two-week stretch: OpenAI cut GPT-5.6 Luna by 80 percent on 30 July, and Anthropic announced on 11 August that Claude Sonnet 5's introductory pricing of $2/$10 is now permanent. All three vendors just repriced the tier where enterprise volume actually lives. If your team runs classification, extraction, summarisation, document pipelines or agentic tooling on Azure, the model you picked for those workloads in the spring is probably no longer the obvious choice, and the spread between the cheapest and the most expensive option is now roughly nine to one on list price.

What Google shipped on 13 August

Gemini 3.7 Flash arrived just three weeks after Gemini 3.6 Flash went GA on 21 July, an unusually short cycle even by 2026 standards. Google's launch post positions it squarely at coding and agentic work, with vendor-reported gains over 3.6 Flash on every benchmark it cites: 43.6% versus 34.4% on FrontierCode 1.1 Main, 65.3% versus 49.0% on DeepSWE v1.1, a WebDev Arena Elo of 1588 versus 1538, 34.0% versus 22.0% on the GDP.pdf knowledge-work benchmark, and 30.4% versus 17.0% on AutomationBench for business workflows. Those are Google's own numbers on Google's chosen benchmarks, so treat them as directional rather than decisive. The model also ships with updated safeguards against misuse in CBRN and cyber-offense domains.

The pricing mechanics deserve attention. The $0.75/$3.75 introductory rate runs through the end of 2026, after which standard pricing of $1.50/$7.50 applies. That is a doubling on 1 January 2027, written into the launch terms. Availability at GA spans Google AI Studio, Android Studio, and Google's enterprise agent platform. What it does not span is Azure: Gemini models are not available through Azure AI Foundry or any Microsoft channel, which matters more for procurement than for engineering, as we cover below.

Two weeks that repriced the whole tier

30 July: OpenAI cuts Luna by 80 percent

GPT-5.6 Luna launched on 9 July 2026 as the fast, cost-efficient member of the GPT-5.6 series, aimed at high-volume, latency-sensitive tasks: chat, classification and lightweight agentic workflows. On 30 July OpenAI cut its price by 80 percent, to $0.20 per million input tokens and $1.20 per million output tokens, with cache reads at $0.02 per million. Luna carries a context window of roughly one million tokens with up to 128K completion tokens and a February 2026 knowledge cutoff.

For Azure teams there is a wrinkle. Microsoft stated that Azure OpenAI customers would receive the same price decreases effective 1 August. As of this writing, customers on Microsoft Q&A report that the Azure pricing pages and their billing still reflect the old rates: $1.10/$6.60 for Luna, which is more than five times the direct OpenAI price on input. If you run Luna through Azure OpenAI, check your invoice line items before assuming the cut has landed, and open a support ticket if your August billing shows pre-cut rates after Microsoft's stated effective date.

11 August: Anthropic locks Sonnet 5 at $2/$10

Claude Sonnet 5 launched on 30 June 2026 at an introductory $2 per million input tokens and $10 per million output tokens, with a scheduled increase to $3/$15 after 31 August. On 11 August Anthropic cancelled that increase and made the introductory pricing permanent. On TechCrunch's reporting of Anthropic's launch figures, Sonnet 5 scores 63.2% on agentic coding, against 69.2% for Opus 4.8 and 58.1% for its predecessor Sonnet 4.6, and Anthropic describes it as approaching Opus-level capability at a fraction of the cost. The pitch is explicitly agents: planning, tool use with browsers and terminals, and autonomous multi-step work.

For Azure teams Sonnet 5 is notable for a second reason: it has been generally available in Microsoft Foundry since 1 July 2026, with Azure billing through the Microsoft Marketplace at standard Anthropic API rates, Entra ID governance, and a 1M-token context window. Microsoft lists availability in 14 global regions under the Global Standard deployment type. That makes Sonnet 5 the first Claude workhorse-tier model a Swedish enterprise can consume entirely inside its existing Azure commercial agreement.

The three contenders side by side

Gemini 3.7 Flash (GA 13 Aug 2026): $0.75/$3.75 per 1M tokens until 31 December 2026, then $1.50/$7.50. Strong vendor-reported coding and agentic benchmarks. Available via the Gemini API and Google's platforms; not available on Azure.
GPT-5.6 Luna (launched 9 Jul, repriced 30 Jul): $0.20/$1.20 per 1M tokens direct from OpenAI, cache reads $0.02. About 1M-token context, 128K max output. On Azure OpenAI, but the price cut had not visibly propagated to Azure pricing pages as of mid-August.
Claude Sonnet 5 (launched 30 Jun, price locked 11 Aug): $2/$10 per 1M tokens, now permanent. 1M-token context on Microsoft Foundry, GA since 1 July with Azure-native billing and Entra ID integration.

On list price alone Luna wins by a wide margin, Gemini 3.7 Flash sits in the middle during its introductory window, and Sonnet 5 is the premium option. But list price is the wrong number to optimise, for two reasons.

List price is not cost per task

First, the arithmetic on a concrete workload. Take a document pipeline processing 50M input tokens and producing 10M output tokens per day. At list prices, a 30-day month costs approximately:

Luna (OpenAI direct): 50 × $0.20 + 10 × $1.20 = $22/day, about $660/month.
Luna (Azure, pre-cut rates): 50 × $1.10 + 10 × $6.60 = $121/day, about $3,630/month. This is why verifying the Azure price propagation matters.
Gemini 3.7 Flash (intro): 50 × $0.75 + 10 × $3.75 = $75/day, about $2,250/month. From January 2027: $150/day, about $4,500/month.
Sonnet 5: 50 × $2 + 10 × $10 = $200/day, about $6,000/month.

Second, and more important: token consumption per completed task differs enormously between models, especially on agentic work. A model that thinks longer, calls more tools, or writes more verbose intermediate output can consume several times the tokens of a terser model on the same task. Reasoning tokens are typically billed at the output rate, which is the expensive rate on every one of these price lists. A model with a higher per-token price that finishes a task in fewer, shorter turns can cost less per task than a nominally cheaper one. Caching changes the picture again: Luna's $0.02 cache reads make repeated-prefix workloads dramatically cheaper if your prompt architecture is cache-friendly. None of this shows up in a pricing table, which is why the only trustworthy number is cost per completed task measured on your own workload.

One more budgeting note: two of these three prices are introductory. Gemini 3.7 Flash doubles on 1 January 2027 by published schedule. Sonnet 5's introductory price was scheduled to rise and then was made permanent, which was a pleasant surprise, but you should not build a 2027 budget on the assumption that vendors will keep repeating that move. Budget at standard rates and treat introductory windows as margin.

A decision guide for Azure-first teams

High-volume, simple tasks (classification, extraction, routing, short summaries): Luna's list price is hard to argue with, and these tasks rarely stress the capability ceiling. If you must stay inside Azure, confirm the Azure OpenAI rate for Luna actually matches OpenAI's cut before committing volume.
Agentic coding and tool-heavy workflows inside Azure: Sonnet 5 on Microsoft Foundry is the path of least procurement resistance: Azure billing, Entra ID, no new vendor contract. Its agentic benchmark positioning and 1M context are aimed at exactly this workload. Measure cost per task, not per token, before ruling it out on price.
Coding and agent workloads where Google Cloud is already in the estate: Gemini 3.7 Flash's benchmark deltas over 3.6 Flash are large, and the introductory price is competitive. For an Azure-only shop, weigh whether the model gains justify a second cloud relationship.
Anything long-lived: model the January 2027 prices, not the August 2026 prices, and keep your prompts and evals portable enough that switching models is a config change rather than a rewrite.

The Swedish and EU angle

Procurement is half the decision. For most Swedish enterprises and public-sector bodies, adding a model that lives inside the existing Microsoft agreement is an approval memo; adding one that requires a new Google Cloud relationship is a procurement project, potentially an upphandling. Luna and Sonnet 5 both bill through Azure. Gemini 3.7 Flash does not, and no Microsoft channel for Gemini exists. That asymmetry can outweigh a benchmark delta, and it is worth stating plainly in any model-selection document so the decision is made consciously rather than by default.

Data residency needs deployment-type reading, not vendor-name reading. "It runs on Azure" is not a residency guarantee. Claude models in Microsoft Foundry under the Global Standard deployment type are documented as served from both Azure and Anthropic infrastructure across global regions, and the documented data-zone option for Claude is a US data zone. Azure OpenAI has offered EU Data Zone deployment types for GPT models, but you must verify that GPT-5.6 Luna is available in that deployment type in your region before promising EU-bounded processing to your DPO. For Gemini, EU processing commitments are a Google Cloud contractual question. In every case, the artifact your compliance function needs is the deployment-type documentation, not the marketing page.

The AI Act makes vendor documentation a selection criterion. The GPAI transparency obligations under the EU AI Act have applied to newly placed models since August 2025, and all three vendors now publish model documentation in some form. If you are building the technical documentation for your own AI systems, the practical question is how easily each vendor's published material maps into your Article 13-style transparency duties and your internal model registry. Make "quality of the model card and data documentation" an explicit line in your evaluation rubric; it is cheap to score and it predicts how painful your audits will be.

Run the eval before you commit

Every benchmark number above is vendor-reported, on benchmarks the vendor selected, and the three vendors do not publish results on a common suite, so cross-vendor comparisons from these figures are not valid. The workhorse tier is where a weekend of evaluation work pays for itself fastest, because the volumes are large enough that a 20 percent cost-per-task difference is real money. Build a set of 100 to 200 representative tasks from your actual workload, run all three models against it through their respective APIs, and record three numbers per model: task success rate against your own rubric, total tokens consumed per completed task, and wall-clock latency. Multiply tokens by the January 2027 list prices. That table, not this article, should make your decision.

1. Inventory your workhorse-tier workloads and their monthly token volumes.
2. Verify what you are actually paying on Azure for GPT-5.6 Luna against OpenAI's 30 July cut.
3. Build a 100-task eval from real workload samples; measure success rate and tokens per completed task on all three models.
4. Budget at standard 2027 rates: $1.50/$7.50 for Gemini 3.7 Flash, and treat every introductory price as temporary.
5. Confirm deployment-type residency guarantees in writing before promising EU-bounded processing.
6. Record the procurement cost of each option, including whether it requires a new cloud vendor relationship.
7. Keep prompts and evals model-portable so the next reprice, which is coming, is a config change.